Search

This paper presents a new image processing and analysis technique for the quality evaluation of cell nuclei to support medical diagnostics in breast cancer. The technique allows cell nuclei that are deformed or overlapped by biological material to be reconstructed. The paper proposes a sensitivity and similarity approach, enriching the PatchMatch correspondence algorithm in accurate cell reconstruction. Its application in reconstruction processes enables accelerated computations and an increased probability of obtaining appropriate segmentation results. The numerical results demonstrate that the developed system allows for automatic and effective cell nuclei reconstruction with an acceptable average area accuracy level above 85% compared with manual human results (assuming manual segmentation as a true value). The reconstruction system allows for the recovery of the proper shape of the analyzed distorted cells very rapidly and in a repeatable manner. An additional advantage of the procedure is that the nuclei area overlapped by artifacts or other cells can be determined. The experimental results prove the high utility of the method in final HER2 gene amplification assessment in breast cancer images.

This research introduces the Recursive General Regression Neural Network Oracle (R-GRNN Oracle) and is demonstrated on several binary classification datasets. The traditional GRNN Oracle classifier (Masters et al., 1998) combines the predictive powers of several machine learning classifiers by weighing the amount of error each classifier has on the final predictions. Each classifier is assigned a weight based on the percentage of errors it contributes to the final predictions as the classifiers evaluate the dataset. The proposed R-GRNN Oracle is an enhancement to the GRNN Oracle in which the proposed algorithm consists of an oracle within an oracle – where the inner oracle acts as a classifier with its own predictions and error contribution. By combining the inner oracle with other classifiers, the R-GRNN Oracle produces superior results. The classifiers considered in this study are: Support Vector Machine (SVM), Multilayer Perceptron (MLP), Probabilistic Neural Network (PNN), Gaussian Naive Bayes (GNB), K-Nearest Neighbor (KNN), and Random Forest (RF). To demonstrate the effectiveness of the proposed approach, several datasets were used, with the primary one being the publicly available Spambase dataset. The predictions of SVM, MLP, KNN, and RF were used to create the first GRNN Oracle, which was then enhanced with the high performances of SVM and RF to create the second oracle, the R-GRNN Oracle. The combined recursive model was 93.24% accurate using 10-fold cross validation, higher than the 91.94% of the inner GRNN Oracle and the 91.29% achieved by RF, the highest performance by a stand-alone classifier. The R-GRNN Oracle was not only the most accurate, but it also had the highest AUC, sensitivity, specificity, precision, and F1-score (97.99%, 91.86%, 94.40%, 93.28%, and 92.57%, respectively). The research contribution of this paper is introducing the concept of recursion (a concept not fully explored in machine learning models and applications) and testing this structure's ability on further enhancing the performance of the traditional oracle. The recursive model has also been applied to several other datasets: The Human Resources, Bank Marketing, and Monoclonal Gammopathy of Undetermined Significance (MGUS) datasets. The results of these implementations are summarized in this paper.

This paper presents a new semantic kernel for classification of high-dimensional data in the framework of Support Vector Machines (SVM). SVMs have gained widespread application due to their relatively higher accuracy. The efficacy of SVMs, however, depends upon the separation of the data itself as well as the kernel function. Text data, for instance, is difficult to classify due to synonymy and polysemy in its contents, having multi-topical instances that can result in mislabeling, and being highly sparse in the bag-of-words representation. While the soft margin parameter and kernel tricks are used in SVM to deal with outliers and non-linearly separable data, using data statistics and correlation has not been fully explored in the literature. This paper explore the use co-similarity (i.e., soft co-clustering) to find latent relationships between documents motivated by the success of co-clustering and subspace clustering methods. It has been shown that the use of weighted higher-order paths between instances in the data can be a good measure of similarity values which can then be used for both classification and to correct mislabeled (or outlier) data in the training set. The proposed kernel is generic in nature and suitable for sparse, dyadic data where direct co-occurrences are not necessary common as in the case of textual data, link-analysis in social media networks, co-authorship, etc. It also studies the impact of noise in the training data and provides a technique to re-label such instances. It is also observed that re-labelling of selected training data reduces the adverse effect of outliers or label noise and can greatly improve the classification of the test data. To the best of our knowledge, we are the first to introduce a supervised co-similarity based kernel function and also provide mathematical formulation to show that it is a valid Mercer's kernel. Our experiments show that the proposed framework outperforms current and state-of-the-art methods in terms of classification accuracy and is more resilient to label noise.

Facial image manipulation is a particular instance of digital image tampering, which is done by compositing a region from one facial image into another facial image. Fake images generated by facial image manipulation now spread like wildfire on news websites and social networks, and are considered the greatest threat to press freedom. Previous research relied heavily on handcrafted features to analyze tampered regions which were inefficient and time-consuming. This paper introduces a framework that accurately detects manipulated face image using deep learning approach. The original contributions of this paper include (1) a customized convolutional neural network model for Manipulated Face (MANFA) identification; it contains several convolutional layers that effectively extract features of multi-levels of abstraction from a tampered region. (2) A hybrid framework (HF-MANFA) that uses Adaptive Boosting (AdaBoost) and eXtreme Gradient Boosting (XGBoost) to deal with the imbalanced dataset challenge. (3) A large manipulated face dataset that is manually collected and validated. The results from various experiments proved that proposed models outperformed existing expert and intelligent systems which were usually used for the manipulated face image detection task in terms of area under the curve (AUC), computational complexity, and robustness against imbalanced datasets. As a result, the presented framework will motivate the finding of a more powerful altered face images detection method and encourages the integration of the proposed model in applications that have to deal with manipulated images regularly.

Identification of tumor in medical images is an important task which leads to increased survival chances. In many cases, tumor diagnosis in medical images is very complicated since the pixel intensities between benign and malignant tissues may be very close. The identification problem is defined as the intensification of the mutual data between the image pixel intensities and the desired areas (tumors in this paper), depending on a limitation of the overall size of the desired area borders. This paper introduces a new approach for image segmentation in differentiating between benign and malignant tissues through a nonparametric approach. The proposed approach assumes that the probability densities correlated with pixel intensities in the image for each region are not already known. The image intensity is used instead of using the probability. A nonparametric density estimation is proposed and formulates the theoretical data optimization problem by applying curve evolution methods and deriving the correlated gradient flows. It uses level-set techniques to achieve the resulting evolution. The algorithm is applied to 100 sets of different real data in patient images with benign or/and malignant tissues. The experimental results show that the proposed algorithm proved effective in tumor identification with an acceptable accuracy of 93.1%, especially in-patient images diagnosed with carcinoma at an early stage.

Deep convolutional models have been able to get extraordinary results in visual, speech and textual processing domains. Nevertheless, the complexity of manifold of images in data space necessitates the deep networks to have a large number of parameters and subsequently these deep models prone to redundancy and overfitting. One of the most effective methods to tackle this problem is devising special regularization methods in the context of convolutional models. In this paper, after presenting and discussing the most important models in the field, sparse feature maps are proposed and employed as a penalty term in the cost function. This conducts the learning process to construct kernels so that the feature maps at every point are sparse along the depth. A sparse representation is able to adapt to the varying level of information of natural images and can help to extract more independent and informative representations. Also, the DropMaps concept is proposed and employed in the last convolutional layer of the model. This technique applies dropout on the feature maps that causes coincidence of feature maps to be avoided. As is shown in this paper, sparse feature maps and DropMaps can handle the problem of overfitting in large models for tiny images. We have studied the effect of the sparsity rate on the accuracy of the model, and it is observed the accuracy of the test dataset reaches its maximum at a sparsity rate of 0.05. Moreover, by designing appropriate learning rate curves, we were able to obtain ensemble machines with much less cost for training. It is noticeable that the test accuracy is higher than the validation accuracy of the ensemble, indicating that the model has not overfitted. In the input of the model, a random online preprocessing layer is employed for the training phase that helps regularization of the model. Comparing input space and feature space of the model we found that the proposed network is able to successfully separate images of different classes. Finally, testing the proposed model with MNIST dataset has shown that the test set can be classified with accuracy 99.75. The same test with CIFAR 10 dataset attained an accuracy of 94.05.

Accurate Magnetic Resonance Imaging (MRI) image segmentation is a clinically challenging task. More often than not, one type of MRI image is insufficient to provide the complete information about a pathological tissue or a visual object from the image. As a result, radiology experts often combine multisequence images of a patient to verify the location, extension, prognosis and diagnosis of an object. There are mainly two challenges in medical image segmentation. One is ambiguous boundary that appears between an object and its neighboring region, and the other is intensity inhomogeneity that appears within a region. Thus, this paper focuses on how to effectively segment multisequence medical images despite these two main challenges. This paper proposes a multi-phase approach that integrates both data and domain knowledge into multisequence MR image segmentation. This study divides the segmentation approach into three phases, which are (i) information modeling, (ii), information fusion, and (iii) visual object extraction. In the first phase, random walks algorithm is modified and used to model the information of an image. Because of the ambiguous boundary and intensity inhomogeneity that appear within an image, extra terms related to homogeneity- and object feature-based components are added into the weighting function of random walks algorithm. In the second phase, weighted averaging method is used to fuse information from the image sequences. Both data information of an image as well as user knowledge are integrated to determine the weights of each sequence for fusion. In the final phase, the concept of information theoretic rough sets (ITRS) is utilized to address the issue of ambiguous boundary that may appear between the visual object and its background for object extraction. The proposed approach is tested on MICCAI brain tumor dataset to extract brain tumor and its performance is compared with other established methods. The experiments show promising results, with an average DICE accuracy of 0.7 and 0.63 for high- and low-grade tumor, respectively. As compared to the other fully- and semi-automatic methods that require training and careful initialization processes, the proposed approach is able to extract the brain tumor with prior knowledge about the image.

The paper addresses the problem of designing Brain-Computer Interfaces. It investigates feature selection methods in regression, applied to ECoG-based motion decoding. The problem is to predict hand trajectories from the voltage time series of cortical activity. A special characteristic of this problem is the inherently multi-way structure of feature description. The feature description resides in spatial-spectra-temporal domain and includes the voltage time series and their spectral representation. Since electrocorticographic data is highly correlated in temporal, spectral and spatial domains, redundancy of the feature space as well as its dimensionality become a major obstacle for robust solution of the regression problem both in multi-way and flat cases. Feature selection reduces dimensionality and increases model robustness. It plays the crucial role in obtaining adequate predictions. The main contribution of this paper is the following. We propose a filtering feature selection for multi-way data. The proposed method extends quadratic programming feature selection (QPFS) approach. QPFS selects a subset of features by solving a quadratic problem. It incorporates estimates of similarity between features and their relevance to the regression problem. QPFS offers an effective way to leverage similarities between features and their importance. Our modification allows to apply QPFS to multi-way data. By taking the multi-way structure of features into account, the proposed modification reduces computational costs of optimization problem in QPFS. Experimental outcomes demonstrate that the proposed modification improves prediction quality of resultant models. The proposed method is model-free and provides interpretable results, which makes it relevant for knowledge extraction and domain analysis.

There are approaches that successfully recognize activities of daily living by using a trained classifier on feature vectors created from binary sensor data. Although these approaches have been successful, there are still open issues such as the evaluation of multiple temporal windows, ensembles of classifiers or unbalanced classes which need to be addressed in order to improve the performance of the real-time activity recognition process. In this paper, we present a methodology for Real-Time Activity Recognition based on the diverse fields of Machine Learning, including Fuzzy Logic and Recurrent Neural Networks. The methodology uses a long-term and short-term representation of binary-sensor activations based on Fuzzy Temporal Windows. The paper proposes an ensemble of activity-based classifiers for the purposes of balanced training, where each classifier in the ensemble is a Long Short-Term Memory. The approach was evaluated using two binary-sensor datasets of daily living activities and benchmarked against previous approaches based on the combination of sensor activation features.

Mobile crowdsensing is a recent model in which a group of mobile users uses their smart devices such as smartphones or wearable devices to cooperatively perform a large-scale sensing task. In this paper, a novel model will be introduced for recognizing/classifying human activities that were collected from sensor units on the chest, legs, and arms. The proposed model employed the k-Nearest Neighbor (k-NN) classifier which is one of the most common classifiers. k-NN has only one parameter, k, to determine the number of selected nearest neighbors to the test or unknown samples for predicting the class labels of the unknown samples. Searching for the value of k which has a great impact on the classification performance is difficult especially with high dimensional data. This paper employs the Particle Swarm Optimization (PSO) algorithm to search for the optimal value of the k parameter in the k-NN Classifier. This paper shows first experimentally how the PSO in the proposed algorithm searches for the optimal value of k parameter to reduce the misclassification rate of the k-NN classifier. Then, in the second experiment, ten standard datasets are utilized to benchmark the performance of the proposed algorithm. For verification, the results of the PSO-kNN algorithm are compared with two well-known algorithms: Genetic Algorithm (GA) and Ant Bee Colony Optimization (ABCO). In the third experiment, the proposed PSO-kNN algorithm was employed for recognizing human activities. The experimental results proved that the PSO-kNN algorithm is able to find the optimal or near optimal value(s) of the k parameter which enhances the accuracy of k-NN classifier. The results also demonstrated lower error rates compared when GA and ABCO algorithms.

Advances in neuroimaging techniques have enabled early diagnosis and better understanding of neuro-psychological disorders using resting state - functional Magnetic Resonance Images (rs-fMRI). However, factors like the large intra-class variability, high homogeneity, limited sample size and varied acquisition methodologies in aggregated datasets have limited their classification performance. These issues especially are highly prevalent in the automatic diagnosis of Attention Deficit Hyperactivity Disorder (ADHD). In this paper, a Discriminative Spatial Filtering Method (DSFM) is developed to improve the separability between the two classes in the dataset by automatically identifying regions of the brain with abnormal activities. Further, DSFM uses an orthogonal projection to extract highly separable features from raw rs-fMRI data of the identified regions of the brain. Using a projection based learning classifier DSFM achieved a classification accuracy of 73.83 % on the publicly available ADHD200 dataset. The results show that, selective inclusion of data from the regions of the brain with discriminant activity identified using DSFM removes the redundant features and improved the classification performance (improvement of 3.5 % over baseline). Further, the brain activity maps derived by inverse mapping of the DSFM spatial filters describe the nature of the abnormal brain activities and identify the corresponding regions responsible for the cognitive symptoms in neuro-psychological disorders like ADHD. The results in this paper clearly show that DSFM is a reliable diagnostic tool with interpretable results critical for the accurate understanding the etiological factors of neuro-psychological disorders like ADHD using rs-fMRI.

Convolutional neural network is a prominent innovation in computer vision but is often troubled by problems such as dark light, turbidity, blur and high similarity to the background when applied to underwater object detection. Underwater object detection is one of the basic techniques of underwater grasping automation which plays a very important role in ocean detection and fishery of aquatic products. This paper presented an automatic detection method of underwater sea cucumber based on deep learning, which will provide effective technical support for the automated breeding and harvesting of sea cucumber. The Shortcut Feature Pyramid Network (S-FPN) proposed in this paper improves the existing multi-scale feature fusion strategy through shortcut connection. The ablation experimental results show that the mean average precision (mAP) of S-FPN reaches 91.5% which outperforms the baseline Feature Pyramid Network (88.6%), YOLO v3 (83.7%) and SVM-HOG (61.6%). To resolve the problem of complex environmental background interference of ocean floor, we proposed a Piecewise Focal Loss (PFL) function for balancing the positive and negative samples such that the algorithm can focus on the training difficulty of hard (i.e., positive) samples. And the ablation experimental results show that the mAP of PFL reaches 92.3% which outperforms the baseline Cross Entropy (91.5%) and Focal Loss (91.8%). Also, we chose Exponential Linear Unit as the optimization strategy, and Adaptive Moment Estimation as the activation function by ablation research, finally the mAP reached 94%.

Over decades, medical image diagnosis has gained prominence, as it has saved millions of lives from dreadful diseases. Improvising image processing techniques in medical images has provoked substantial increment in the efficiency levels of patient diagnosis. Segregation and visualization of pathologies have been made available by several image processing algorithms, under which tumor recognition and tissue segmentation in Magnetic Resonance (MR) brain images are accomplished in this paper using a novel approach. The novel methodology suggested through this paper ensemble the functioning of two different techniques well known to the research community. The techniques are Bacteria Foraging Optimization (BFO) and Modified Fuzzy C – Means (MFCM) algorithms, where, BFO is familiar for its optimization abilities and MFCM, an advancement of Fuzzy C – Means algorithm initiates the clustering operation. Both these techniques are well employed into a single framework to perform MR brain image segmentation, so that effective tumor segregation and tissue segmentation can be achieved, concurrently. Frequent parameter tuning is not required in the case of the proposed combinational algorithm, which is entirely an automated approach, and development of such algorithm would facilitate the radiologists in patient diagnosing procedures, as they extricate both manual intervention and large time consumption. With the support rendered by an automated algorithm, large volumes of clinical datasets could be assessed with ease. The proposed algorithm is validated by the radiologists and also using the comparison parameters such as sensitivity, Specificity, Mean Squared Error (MSE), Peak Signal to Noise Ratio (PSNR) and computational time. By analysing the above said parameters it has been proved that the proposed algorithm is prodigious than the state-of-art segmentation algorithms. The sensitivity and the specificity values offered by the proposed methodology are 0.9048 and 0.9825, respectively. In addition to the clinical datasets, MR brain images from Harvard Brainweb database, Brainweb simulated database and BRATS-2013 challenge database are used to demonstrate the segmentation efficiency of the proposed algorithm. These factors are good enough to prove that the proposed combinational framework can be preferably opted for medical image analysis.

Several biological research studies have shown that the number of individuals of certain species of anurans in a specific geographical region, and the evolution of this number over time, can be used as an indicator of climate change. To detect the presence of anurans, Wireless Sensor Networks (WSNs) are usually deployed with the aim of obtaining bio-acoustic information in a set covering numerous locations. However, the identification of the anuran species from a huge number of recordings usually involves an overwhelming task that has to be undertaken by expert and intelligent systems. Previous studies into this issue have proposed several classification techniques with a common approach: they all take into account the sequential characteristic of sounds by considering syllables or other kinds of vocal segments. In noisy sounds, as it is usually the case in recordings made in natural habitats, segmentation of the signal is no straightforward task and may cause low classification accuracy. To override this problem, a new non-sequential approach is proposed in this paper. It is based on considering very small pieces of sounds (frames) each of which is then classified without considering preceding or subsequent information. Up to nine frame-based classifiers are explored in this paper and their performances are compared to the most commonly used sequential classifier: the Hidden Markov Model (HMM). Additionally, for featuring the frames, many choices have been described, although the application of the Mel Frequency Cepstral Coefficients (MFCCs) has probably become the most common method. In this work, an alternative methodology is suggested: the use of a set of MPEG-7 parameters, which offers a normalized solution with a much greater semantic content. The experimental results have shown that the proposed method clearly outperforms the HMM, thereby showing the non-sequential classification of anuran sounds to be feasible. From among the algorithms tested, the decision-tree classifier has shown the best performance with an overall classification success rate of 87.30%, which is an especially striking result considering that the analyzed sounds were affected by a decidedly noisy background.

In computer vision, significant advances have been made on object detection with the rapid development of deep convolutional neural networks (CNN). This paper provides a comprehensive review of recently developed deep learning methods for small object detection. We summarize challenges and solutions of small object detection, and present major deep learning techniques, including fusing feature maps, adding context information, balancing foreground-background examples, and creating sufficient positive examples. We discuss related techniques developed in four research areas, including generic object detection, face detection, object detection in aerial imagery, and segmentation. In addition, this paper compares the performances of several leading deep learning methods for small object detection, including YOLOv3, Faster R-CNN, and SSD, based on three large benchmark datasets of small objects. Our experimental results show that while the detection accuracy on small objects by these deep learning methods was low, less than 0.4, Faster R-CNN performed the best, while YOLOv3 was a close second.

This paper introduces a three-step framework for classifying multiclass radiography images. The first step utilizes a de-noising technique based on wavelet transform (WT) and the statistical Kolmogorov Smirnov (KS) test to remove noise and insignificant features of the images. An unsupervised deep belief network (DBN) is designed for learning the unlabelled features in the second step. Although small-scale DBNs have demonstrated significant potential, the computational cost of training the restricted Boltzmann machine is a major issue when scaling to large networks. Moreover, noise in radiography images can cause a significant corruption of information that hinders the performance of DBNs. The combination of WT and KS test in the first step helps improve performance of DBNs. Discriminative feature subsets obtained in the first two steps serve as inputs into classifiers in the third step for evaluations. Five frequently used classifiers including naive Bayes, radial basis function network, random forest, sequential minimal optimization, and support vector machine and four different case studies are implemented for experiments using the Image Retrieval in Medical Application data set. The experimental results show that the three-step framework has significantly reduced computational cost and yielded a great performance for multiclass radiography image classification. Along with effective applications in image processing in other fields published in the literature, deep learning network in this paper has again demonstrated its robustness in handling a complex set of medical images. This implies that the proposed approach can be implemented in real practice for analysing noisy radiography images, which have many useful medical applications such as diagnosis of diseases related to lung, breast, musculoskeletal or pediatric studies.

A facial beauty analysis toward attractiveness assessment application is presented.A geometric facial beauty score function is proposed for facial aesthetic perceptron.A semi-supervised learning with Hessian graph and random projection is proposed.A novel geometric facial beauty (GFB) database is provided in this paper. Facial beauty analysis has been an emerging subject of multimedia and biometrics. This paper aims at exploring the essence of facial beauty from the viewpoint of geometric characteristic toward an interactive attractiveness assessment (IAA) application. As a result, a geometric facial beauty analysis method is proposed from the perspective of machine learning. Due to the troublesome and subjective beauty labeling, the accurately labeled data scarcity is caused, and result in very few labeled data. Additionally, facial beauty is related to several typical features such as texture, color, etc., which, however, can be easily deformed by make-up. For addressing these issues, a semi-supervised facial beauty analysis framework that is characterized by feeding geometric feature into the intelligent attractiveness assessment system is proposed. For experimental study, we have established a geometric facial beauty (GFB) dataset including Asian male and female faces. Moreover, an existing multi-modal beauty (M2B) database including western and eastern female faces is also tested. Experiments demonstrate the effectiveness of the proposed method. Some new perspectives on the essence of beauty and the topic of facial aesthetic are revealed. The impact of this work lies in that it will attract more researchers in related areas for beauty exploration by using intelligent algorithms. Also, the significance lies in that it should well promote the diversity of expert and intelligent systems in addressing such challenging facial aesthetic perception and rating issue.

Solving a real-world problem with an uncommon dataset.Highly time-efficient and accurate for classification of high resolution X-Ray images.Emphasizing on keypoints of the image instead of considering all parts of the image.Considering the dependency between the visual words in the bag of visual words.Adopting Tree-Augmented Bayes in the task of image classification. Smuggling has long played an important role in the inefficiency of economies. To secure the borders against this illegal act, X-Ray Inspection Systems are often deployed at the borders and customs. In this paper, we present a new method for classification of shipping containers X-Ray images, produced in the inspection lines. The aim is to improve the matching accuracy of the presented manifest, which lists the claimed contents of the shipping containers, with the real contents of the container. The proposed method is based on utilizing Scale Invariant Feature Transforms (SIFT) feature vectors, Bag of visual words (BOVW) and tree augmented naive Bayes (TAN) approach for classifying containers X-Ray images. The prior research on classification of X-Ray images of shipping containers has focused mostly on working with greedy algorithms such as sliding windows for task of classification. More recent studies introduced filter banks and visual words for extraction of features. The proposed method for the first time considers the salient points and keypoints for the task of feature extraction. In addition, this paper presents a framework using the tree augmented naive Bayes based on the theory of learning Bayesian networks, which is proved to have a significant improvements upon the prior designed systems by considering the correlations among the extracted features. For experimental evaluations, our method is compared with two recently proposed methods on containers X-Ray images categorization. The results show that the proposed method is more accurate and time-efficient in categorization of X-Ray images.

A new initialization technique is proposed to improve the performance of K-means.A data transformation approach is proposed to solve empty cluster problem.An efficient method is proposed to estimate the optimal number of clusters.Proposed clustering method provides more accurate clustering results. Clustering provides a knowledge acquisition method for intelligent systems. This paper proposes a novel data-clustering algorithm, by combining a new initialization technique, K-means algorithm and a new gradual data transformation approach to provide more accurate clustering results than the K-means algorithm and its variants by increasing the clusters coherence. The proposed data transformation approach solves the problem of generating empty clusters, which frequently occurs for other clustering algorithms. An efficient method based on the principal component transformation and a modified silhouette algorithm is also proposed in this paper to determine the number of clusters. Several different data sets are used to evaluate the efficacy of the proposed method to deal with the empty cluster generation problem and its accuracy and computational performance in comparison with other K-means based initialization techniques and clustering methods. The developed estimation method for determining the number of clusters is also evaluated and compared with other estimation algorithms. Significances of the proposed method include addressing the limitations of the K-means based clustering and improving the accuracy of clustering as an important method in the field of data mining and expert systems. Application of the proposed method for the knowledge acquisition in time series data such as wind, solar, electric load and stock market provides a pre-processing tool to select the most appropriate data to feed in neural networks or other estimators in use for forecasting such time series. In addition, utilization of the knowledge discovered by the proposed K-means clustering to develop rule based expert systems is one of the main impacts of the proposed method.

This paper describes a new hybrid approach, based on modular artificial neural networks with fuzzy logic integration, for the diagnosis of pulmonary diseases such as pneumonia and lung nodules. In particular, the proposed approach analyzes medical images, which are digitized chest X-rays, focusing on a classification method based on descriptors, such as grayscale histogram features, gray-level co-occurrence matrix (GLCM) texture-based features, and local binary pattern texture features. Then, to perform feature reduction, a multi-objective genetic algorithm is used to obtain an optimized neuro-fuzzy classifier, which is able to classify the pathology found in the analyzed chest X-ray. The main contribution of this paper is the proposed modular neural network approach, which divides features to achieve specialized analysis in the modules for digital image analysis and classification. The proposed approach achieves high classification accuracy after evaluating the neuro-fuzzy model with three large datasets of chest X-rays.

This paper provides a comprehensive survey of various techniques used in computer vision for the automatic identification of plants with the help of leaf images. The extracted information is used by the botanists to identify different species of plants and use their medicinal or other properties. With the upsurge of human interference, the number of plants seems to decrease, but their automatic identification can lead to conservation. The Leaf images may be acquired by a phone camera or a digital camera mounted on a tripod stand. The leaves may be covered by dirt, shadows, or hidden under other leaves. Real-life applications based on the automatic identification of plants can successfully identify even similar-looking plant leaves in all environmental conditions. This paper provides a state-of-the-art review of different leaf extraction techniques which are categorized according to the features of leaf used and their pros and cons. We also discuss and compare the various classifiers used in the identification process. The conclusion of the paper also provides different areas of improvement and future work.

Feature selection is crucial to develop a brain computer interface (BCI) system which has high classification accuracy and less computational complexity in especially a large feature space. Feature selection (FS) problem has been solved by many various methods. Among these methods, especially evolutionary computation (EC) techniques have gained a lot of attention in recent years. However, there are very few studies in the literature that consider FS problem as a multi-objective problem to find the optimal trade-off between classification accuracy and the number of selected features. Therefore, in this paper, a non-dominated sorting multi-objective symbiotic organism search (NSMOSOS) algorithm is proposed to generate the optimal feature subset in BCI. The efficiency and robustness of the proposed algorithm as a feature selection method is investigated in two datasets based on motor imagery. The highest classification accuracies of NSMOSOS for dataset 1 and dataset 2 are obtained 97.86% with 11 features and 96.57% with average 19 features, respectively. The obtained results demonstrate that the proposed method achieves satisfying results with regard to both the classification accuracy improvement and feature reduction rates for both datasets. The superiority of the proposed method is verified compared with the existing methods for both datasets. Besides, three different versions of symbiotic search organism (SOS) algorithm are improved, and pros and cons of these algorithms are evaluated compared with each other. In conclusion, the paper indicates that the proposed NSMOSOS algorithm is an efficient and practicable technique for FS problem and could be helpful in developing the BCI applications.

Random forests are grown on convolutional neural networks for scene categorization.Features from multi-layers of deep convolutional neural networks are utilized.A feature selection method is proposed to use random forests to categorize scenes. Breakthrough performances have been achieved in computer vision by utilizing deep neural networks. In this paper we propose to use random forest to classify image representations obtained by concatenating multiple layers of learned features of deep convolutional neural networks for scene classification. Specifically, we first use deep convolutional neural networks pre-trained on the large-scale image database Places to extract features from scene images. Then, we concatenate multiple layers of features of the deep neural networks as image representations. After that, we use random forest as the classifier for scene classification. Moreover, to reduce feature redundancy in image representations we derived a novel feature selection method for selecting features that are suitable for random forest classification. Extensive experiments are conducted on two benchmark datasets, i.e. MIT-Indoor and UIUC-Sports. Obtained results demonstrated the effectiveness of the proposed method. The contributions of the paper are as follows. First, by extracting multiple layers of deep neural networks, we can explore more information of image contents for determining their categories. Second, we proposed a novel feature selection method that can be used to reduce redundancy in features obtained by deep neural networks for classification based on random forest. In particular, since deep learning methods can be used to augment expert systems by having the systems essentially training themselves, and the proposed framework is general, which can be easily extended to other intelligent systems that utilize deep learning methods, the proposed method provide a potential way for improving performances of other expert and intelligent systems.

Because the most dangerous type of skin cancer, melanoma, is very difficult for dermatologists to detect because of the low contrast between the lesion and the adjacent skin, the automatic application of skin lesion segmentation is regarded as very challenging. This paper proposes the implementation of a medical image segmentation that will accelerate a melanoma diagnosis by dermatologists. In the implementation, Fully Convolutional Network (FCN) architectures generated by modifying Convolutional Neural Network (CNN) architectures are used. The proposed algorithm for an automatic semantic segmentation of skin lesions utilizes four different FCN architectures, FCN-AlexNet, FCN-8s, FCN-16s, and FCN-32s. The experimental studies in this paper are constructed on the ISIC 2017 dataset, and the evaluations of these architectures on the dataset are carried out for the first time with this study. In the experimental studies, once the images in the dataset are preprocessed, the FCNs are first trained separately. Secondly, the accuracies and Dice coefficients on the validation dataset are calculated by using these trained FCN architectures. Thirdly, the obtained results are compared. Finally, the inferences of lesion segmentation are visualized in order to exhibit how exactly the FCN architectures can segment the lesions. The experimental results show that the FCNs in the proposed algorithm are suitable for skin lesion segmentation. In addition, it is thought that the experimental results will contribute to the scientific literature and assist the researchers who are working on medical image segmentation.

Adaptive learning rate has been proposed for Deep Learning in MLP.Technique for updating learning rate is free of hyper-parameters.Learning rate is a function of a parameter called learning parameter.Learning parameter is updated based on error gradient.Learning rate is further updated based on the Laplacian score of activation values. An attempt has been made to improve the performance of Deep Learning with Multilayer Perceptron (MLP). Tuning the learning rate or finding an optimum learning rate in MLP is a major challenge. Depending on the value of the learning rate, classification accuracy can vary drastically. This issue has been taken as a challenge in this paper. In this paper, a new approach has been proposed to combine adaptive learning rate in conjunction with the concept of Laplacian score for varying the weights. Learning rate is taken as a function of parameter which itself is updated on the basis of error gradient by forming mini-batches. Laplacian score of the neuron isfurther used for updating the incoming weights. This removes the bottleneck involved in finding the optimum value for the learning rate in Deep Learning by using MLP. It is observed on benchmark datasets that this approach leads to increase in classification accuracy as compared to the existing benchmark levels achieved by the well known methods of deep learning.

We presented system for counting the gene markers in FISH images of breast cancer.We have developed the fusion system of several methods.The ensemble of methods improves greatly sensitivity and specificity of system. The paper presents an application of several methods of markers localization in fluorescent in situ hybridization images in order to determine HER2 status of the breast cancer samples. The applied methods implement different mathematical foundations of image processing and perform the analysis in an independent way. The paper proposes their fusion to obtain better results of spot recognition. The integration process of many methods takes into account different points of view on the classification problem and takes the best from each of them in developing the final solution. Thanks to this the improvement of the performance of the system is achieved. The efficiency of the developed system in the form of sensitivity and specificity has been verified on the examples of analysis of many FISH images. They confirm the advantage of using several methods of image processing combined in the form of an ensemble.

Quantification and objective estimation of driver fatigue in real prolonged driving.Simultaneous recording of physiological parameters in wireless and nonintrusive way.Develop a dynamic fatigue detection model by multiple features and contexts. Driver's states in successive time slices are not independent, especially, fatigue is one of a cognitive state that is developing over time. Meanwhile, driver fatigue is also influenced by some corresponding contextual information at a certain time. In such case, classifying driving state at each time slice separately from it in before and after time slices obviously has less meaning. Therefore, a dynamic fatigue detection model based on Hidden Markov Model (HMM) is proposed in this paper. Driver fatigue can be estimated by this model in a probabilistic way using various physiological and contextual information. Electroencephalogram (EEG), Electromyogram (EMG), and respiration signals were simultaneously recorded by wearable sensors and sent to computer by Bluetooth during the real driving. From these physiological information, fatigue likelihood can be achieved using kernel distribution estimate at different time sections. Contextual information offered by specific environmental factors were used as prior of fatigue. As time proceeds, the posterior of fatigue can be gotten dynamically by this HMM-based fatigue recognition method. Based on the results of the method in this paper, it shows that it provides an effective way in detecting driver fatigue.

Developing a new EEG sleep stages classification method based on the statistical features in time domain and complex networks properties.The method provides better EEG sleep signals classification compared with the existing approaches reported.Finding of not all sleep stages can be classified with the same number of the statistical features.Stage Awake can be classified with a fewer statistical features due to including high activity compared with other sleep stages. Sleep stage scoring is a challenging task. Most of existing sleep stage classification approaches rely on analysing electroencephalography (EEG) signals in time or frequency domain. A novel technique for EEG sleep stages classification is proposed in this paper. The statistical features and the similarities of complex networks are used to classify single channel EEG signals into six sleep stages. Firstly, each EEG segment of 30s is divided into 75 sub-segments, and then different statistical features are extracted from each sub-segment. In this paper, feature extraction is important to reduce dimensionality of EEG data and the processing time in classification stage. Secondly, each vector of the extracted features, which represents one EEG segment, is transferred into a complex network. Thirdly, the similarity properties of the complex networks are extracted and classified into one of the six sleep stages using a k-means classifier. For further investigation, in the statistical features extraction phase two statistical features sets are tested and ranked based on the performance of the complex networks. To investigate the classification ability of complex networks combined with k-means, the extracted statistical features were also forwarded to a k-means and a support vector machine (SVM) for comparison. We also compare the proposed method with other existing methods in the literature. The experimental results show that the proposed method attains better classification results and a reasonable execution time compared with the SVM, k-means and the other existing methods. The research results in this paper indicate that the proposed method can assist neurologists and sleep specialists in diagnosing and monitoring sleep disorders.

This paper shows a system for feature optimization using modified Taguchi method.This method can reduce the number of features, and classification hazards.This study enhances the rules of Taguchi method for the system has no exact output.This study classifies multiple clusters with less number of parameters.This method possesses minimal pre-processing with anchor point feature selection. Development of an expert system for clinical application includes automation in diagnosis of abnormality and patient monitoring based on features derived from continuous data set. This paper presents a novel method for feature optimization and classification of electrocardiogram (ECG) for arrhythmia analysis. A feature set optimization technique can reduce the classification hazard by selecting few comprehensive features to cater all kind of abnormalities under consideration. Proposed work deals with ranking and selection of an optimized pair of features using Taguchi method from eleven possible features normally used for characterizing arrhythmic beats like left bundle branch (LBBB), right bundle branch (RBBB) and premature ventricular contraction (PVC) are compared to normal beats. An imposed target based modification of Taguchi method is also suggested for the systems where the output is not pre-defined as in the case of biomedical applications. The proposed method is advantageous for the expert systems in which individual identity of the features are to be stored while reducing the dimensionality of the feature set. Multiclass Navis Bayes classifier is used to classify the beats in a single run and good performance parameters are obtained as reported in the result section.

In this paper, we describe Pixel Sampling Clustering Technique (PSCT), a data-driven sampling procedure used to reduce pixel sets. We view the pixels in an image as a high redundancy 3D space. We also refer to this space as our color model. Our method aims to retain a relevant sample of the data so it can act as a new smaller, hence more efficient, color model. PSCT uses a pair of fast density-based clustering algorithms in tandem. First, it applies Birch and then DBSCAN to keep the most densely represented colors. We cluster the resulting color model and use the labels to segment images. We also complement the sampling method with a refinement algorithm intended to improve color representation. In our paper, we show how to reconstruct images using our reduced color model. We also show that reconstructed images have enough information to perform image related learning tasks with almost the same accuracy than the original images but with only a small fraction of the data. We test our sampling method in three image related supervised and unsupervised tasks and compare them with state-of-the-art methods. For our experiments, we use two image datasets: MIT’s Vision Texture Dataset and Berkeley’s BSD500.

The existing clustering algorithms based on auto-encoder only use one layer of information. In this paper, we propose a new subspace clustering algorithm that uses multiple hidden layers of information from the stacked auto-encoder to construct different kernels. The proposed fuzzy multi-kernel clustering method based on auto-encoder is realized by updating the membership matrix and coefficients of these kernels until the value of objective function is iterated to the minimum error. Also, the proposed method combining auto-encoder achieves high dimension reduction of input data. In order to test the effectiveness of the algorithm, we first performed experiments on the published brain network dataset. Compared with MKFC, RMKKM and other algorithms, the proposed method can significantly improve accuracy. Specifically, the brain network has large dimensions when the functional magnetic resonance imaging (fMRI) data is preprocessed. Therefore, the method in this paper is applied to the Autism Spectrum Disorder (ASD) of the Autism Brain Imaging Data Exchange (ABIDE) database. The experiment results on the network dataset we constructed with high dimension are better than the present several clustering algorithms. The results show that the subspace information after dimensionality reduction is more conducive to clustering.

In the theory of Intuitionistic fuzzy set (IFS), similarity measure is an effective instrument to measure similarity between IFSs. In this paper, we mention limitations of various existing similarity measures and propose a novel dice similarity measure for IFSs. Proposed dice similarity measure is based on inner product and overcomes limitations of existing similarity measures. In order to show the suitability and applicability of proposed dice similarity measure, we implement it on various classification problems of pattern recognition and medical diagnosis. Experimental results show that this does not only overcome limitations of existing similarity measures but also outperforms in pattern recognition and medical diagnosis problems. In this paper, we also propose an algorithm for face recognition problem using proposed dice similarity measure. This algorithm is demonstrated by an example and performance is compared with few existing methods for face recognition problems.

This paper proposes a method to improve neural network classification performance.A linear model was used as post processing of neural network.The parameters of linear model was estimated using Kalman filter iteration.The method can be applied to classify an object regardless of the type of feature.The method has been validated with five different datasets. Neural network has been applied in several classification problems such as in medical diagnosis, handwriting recognition, and product inspection, with a good classification performance. The performance of a neural network is characterized by the neural network's structure, transfer function, and learning algorithm. However, a neural network classifier tends to be weak if it uses an inappropriate structure. The neural network's structure depends on the complexity of the relationship between the input and the output. There are no exact rules that can be used to determine the neural network's structure. Therefore, studies in improving neural network classification performance without changing the neural network's structure is a challenging issue. This paper proposes a method to improve neural network classification performance by constructing a linear model based on the Kalman filter as a post processing. The linear model transforms the predicted output of the neural network to a value close to the desired output by using the linear combination of the object features and the predicted output. This simple transformation will reduce the error of neural network and improve classification performance. The Kalman filter iteration is used to estimate the parameters of the linear model. Five datasets from various domains with various characteristics, such as attribute types, the number of attributes, the number of samples, and the number of classes, were used for empirical validation. The validation results show that the linear model based on the Kalman filter can improve the performance of the original neural network.

Wrinkled fingerprint recognition has been a challenging problem because of changing the position of fingertip features. This change significantly degrades the fingerprint recognition accuracy. Contactless dry three-dimensional (3D) fingerprints have the advantages of reducing the position change of fingertip features presented in both contact-based and contactless dry 2D fingerprints. Unfortunately, in contrast to the contactless dry 3D fingerprints, the position of features in the contactless wrinkled 3D fingerprints will be changed. Furthermore, identifying a fingerprint in a voluminous database is another challenge. With an increasing the number of individuals and inserting their fingerprints into the enrollment database, the cost of identification will increase and can become critical. Fingerprint indexing is a prominent method to reduce the response time of a probe in a large-scale database. The indexing approaches powerfully boost the recognition efficiency of dry fingerprints, but the accuracy of wrinkled fingerprints cannot be guaranteed. Moreover, previous indexing approaches only focused on dry 2D fingerprints and did not consider 3D fingerprints and the problems of wrinkled fingerprints. This paper proposes a 3D fingerprint reconstruction technique based on multi-view contactless wrinkled fingerprint images. In the proposed system, we use two cameras to acquire the frontal image. A dual camera can get more details of an image and is useful to acquire wrinkled fingerprints. In this paper, we propose a rectification technique for wrinkled 2D and 3D fingerprints as well. This paper also proposes a wrinkled fingerprint indexing approach to overcome the problems of wrinkled fingerprints. Our proposal employs minutiae quadruplets, ellipse properties, and a k-means clustering to index and retrieve fingerprints. Finally, a voting technique based on minutiae quality is proposed. Experimental results validate our approach and demonstrate the effectiveness of proposed method. Some new outlooks on the topics of 3D fingerprint acquisition, rectification and 3D reconstruction of wrinkled fingerprints, and fingerprint indexing are revealed. The main impact of this work is that more researchers will be attracted in related areas for wrinkled fingerprint recognition. The main significance is that the diversity of expert systems should be well promoted in addressing reconstruction of 3D shape and rectification of deformed fingerprints.

Human action recognition is defined as determining the actions of humans happening in video sequences. Human action recognition is one of the interesting topics which can play important role in intelligence, surveillance and health protection systems' performance; the performance of human action recognition methods is a vital issue that has been focused on many recent papers. Subspace learning and classification steps can impress the performance of the human action recognition methods. Accordingly, in this paper, new subspace learning and classification algorithms are proposed for human action recognition. Notably, graphs play important role to describe the relationship of data but most of the graph based methods used Euclidean distance metric. To overcome it, a geodesic distance based weighted LASSO l1-graph (GDWL-l1 graph) is proposed to extract the between-class and within-class graphs. Then, double discriminative sparsity preserving projections (DDSPP) algorithm is introduced to map the high-dimensional data to a new discriminant low-dimensional space using these graphs in order to have better discrimination besides sparsity and locality preserving in the mapped space. Subspace learning using DDSPP algorithm leads to a discriminative and sparse low-dimensional space. At the end, a discriminant ridge-based classifier (DRC) is introduced to inherit the grouping effect of the ridge regression besides incorporating the geodesic distance by defining a criterion of the classification. Experimental results show the suitable performance of the proposed method on HMDB51 and UCF101 datasets which are as 66.41% and 92.46%, respectively.

[EN] This paper presents a support vector machine classifier for broken bar detection in electrical induction machine. It is a reliable online method, which has high robustness to load variations and changing operating conditions. The phase current is only physical value to be measured. The steady state current is analyzed for broken bar fault via motor current signature analysis technique based on Hilbert transform. A two dimensional feature space is proposed. The features are: magnitude and frequency of characteristic peak extracted from spectrum of Hilbert transform series of the phase current. For classification task support vector machine is used due to its good robustness and generalization performances. A comparative analysis of linear, Gaussian and quadratic kernel function versus error rate and number of support vectors is done. The proposed classifier successfully detects a broken bar in various operational situations. The proposed method is sufficiently accurate, fast, and robust to load changes, which makes it suitable for use in real-time online applications in industrial drives., This paper is produced as a result of work on FP7 project for area of information and communication technologies named "PRODI - Power plants Robustification based on fault Detection and Isolation algorithms" contract number 224233 financed from European Comity, General Manager for Information community and Media.

Anomaly detection using KPIs (Key Performance Indicators) is a key part of AIOps (Artificial Intelligence for IT Operations). Recent anomaly detection approaches have adopted Machine Learning to detect anomalies on the perspective of individual time points more than events. These approaches do not effectively utilize the labels of continuous anomaly intervals, nor do they pay attention to the differences among anomaly points. The detection performances are therefore not precise enough to be applied in practice, and the differences in length of anomaly intervals also cause loss of performance. In this paper, we propose an anomaly detection approach named ALSR, which uses a label screening model and a relearning model to analyze and utilize the continuous anomaly intervals of KPIs in finer granularity. The label screening model takes advantage of the continuity of anomaly intervals to remove some unnecessary data from the training set, making it more suitable for interval-oriented anomaly detection. The relearning model based on random forest reclassifies the true/false positive points within domain of detected anomalies, thus effectively reduces the number of false positive points. ALSR uses several features extracted by sliding windows, and the feature set is proved to better describe the characteristics of KPI time series. Finally, we conduct comprehensive experiments on 25 KPIs. The total F-score of ALSR is 0.965, which outperforms state-of-the-art anomaly detection approaches.

Parkinson’s disease (PD) is a serious neurodegenerative disorder. It is reported that more than 90% of PD patients have voice impairments. Multiple types of voice recordings have been used for PD detection. Previous work indicates that the use of multiple types of samples per subject degenerates PD detection accuracy. In this paper, we validate it, and propose a two dimensional data selection method for sample and feature selection. The proposed method ranks features by using chi-square statistical model, searches optimal subset of the ranked features and iteratively selects samples. Experimental results show that the proposed method outperforms the state-of-the-art methods in terms of PD detection accuracy on multiple types of voice data.

Low resolution, difficult illumination and noise are the important factors that affect the performance of face recognition system. In order to counteract these adverse factors, in this paper we propose copula probability models based on Gabor wavelets for face recognition. Gabor wavelets have robust performance under lighting and noise conditions. The strong dependencies exist in the domain of Gabor wavelets due to their non-orthogonal property. In the light of the structure characteristic of Gabor wavelet subbands, the proposed methods use copula to capture the dependencies to represent the face image. Three probability-model-based methods CF-GW (Copula Function of Gabor Wavelets), LCM-GW (Lightweight Copula Model of Gabor Wavelets) and LCM-GW-PSO (Lightweight Copula Model of Gabor Wavelets with Particle Swarm Optimization) are proposed for face recognition. Experiments of face recognition show our proposed methods are more robust under the conditions of low resolution, lighting and noise than the popular methods such as the LBP-based methods and other Gabor-based methods. The face features extracted by our methods belong to the Riemannian manifold which is different to Euclidean space. In order to deal the issue of face recognition in complex environment, we can combine the face features in Riemannian manifold with the face features in Euclidean space to obtain the more robust face recognition system by using expert system technologies such as reasoning model and multi-classifier fusion.

This paper presents a max-margin multi-scale convolutional factor analysis (MMCFA) model, which explores the strongly discriminative principle of max-margin learning to improve the classification performance of multi-scale convolutional factor analysis (CFA) model with application to image data classification. Compared with the traditional factor analysis (FA) model, the CFA model can maintain the spatial correlation among the image pixels in two-dimensional space and capture the structural information from images via convolution kernels. Moreover, to extract multi-level features, multi-scale convolution kernels are adopted to capture richer features at different scales of images. Since the unsupervised model may not offer discriminative factors for the classification task, it is expected to introduce the supervised information to the multi-scale CFA model when supervised information is available. To deal with it, a latent variable support vector machine (LVSVM) is linked to the factors learned from multi-scale CFA model, yielding max-margin discrimination, as the classification criterion in the feature space in our proposed model. The multi-scale CFA model and LVSVM learn parameters jointly in a united framework via the Gibbs inference. Experimental results on mixed national institute of standards and technology (MNIST) dataset, Fashion-MNIST dataset, SVHN dataset and measured synthetic aperture radar (SAR) images show that the learned convolution kernels and factors can describe data information well and the proposed model has excellent classification performance.

Noisy image segmentation is a hot topic in natural, medical, and remote sensing image processing. It is among the non-trivial problems of computer vision having to address denoising and segmentation at the same time. Fuzzy C-means (FCM) is a clustering algorithm that has been shown to be effective at dealing with both segmentation-oriented denoising and segmentation at the same time. Moreover, with a high level of noise and other imaging artifacts, FCM loses its ability to perform image segmentation effectively. This paper introduces a Particle Swarm Optimization (PSO)-based feature enhancement approach in the wavelet domain for noisy image segmentation. This approach applies adaptive wavelet shrinkage using FCM clustering performance as an evaluation mechanism and also as the segmentation algorithm. The PSO-based process helps to enhance intensity features for clustering-based denoising, and also provides adaptivity for the system that performs well on a range of real, synthetic, and simulated noisy images with different noise levels and range/spatial properties. Furthermore, the algorithm applies edge enhancement based on Canny edge detector in order to further improve accuracy. Experiments are presented using three different datasets each degraded with different types of common noise. The presented algorithms show effective and consistent performance over a range of severe noise levels without the need for any parameter tuning.

Dynamic texture analysis has been the focus of intensive research in recent years. Thus, this paper presents an innovative and highly discriminative dynamic texture analysis method, whose signature is composed of the weights of the output layer of a randomized neural network after a training procedure. This training is performed by using the pixels of slices of each orthogonal plane of the video (XY, YT, and XT) as input feature vectors and corresponding output labels. The obtained video signature provided an accuracy of 97.05%, 98.54%, 97.74% and 96.51% on the UCLA-50 classes, UCLA-9 classes, UCLA-8 classes and Dyntex++, respectively. These results, when compared to other dynamic texture analysis methods, demonstrate that our descriptors are very effective and that our proposed approach can contribute significantly to the field of dynamic texture analysis.

Classification based on image set is recently a competitive technique, where each set contains multiple images of a person or an object. As a widely used model, affine hull has shown its power in modeling image set. However, due to the existence of noise and outliers, the over-large affine hull usually matches fails when two hulls overlapped. Aiming at alleviating this handicap, this paper proposes a novel method for image set classification, namely Learning of Reduced Prototypes and Local Metric (LRPLM). Specifically, for each gallery image set, a reduced set of prototypes and an optimal local feature-wise metric are simultaneously learned, which jointly minimize the loss function involved the estimation of classification error probability. In doing so, LRPLM inherits the merits of affine hull with better representation to account for the unseen appearances and makes use of the powerful discriminative ability improved by the local metric. It looks like that LRPLM pulls similar image sets with the same class label “closer” to each other, while pushing dissimilar ones “far away”. Extensive experiments illustrate the considerable effectiveness of LRPLM on three widely used datasets. As we know, classification is a research hotspot in expert and intelligent systems. Different from the previous classification methods, LRPLM focuses on image set-based classification technology, while most of them are single-shot classification technology. Thus, the proposed method can be considered as an expert system technology for medical diagnosis, security monitoring, object categorization, and biometrics recognition applications.

A class imbalance problem occurs when a particular class of data is significantly more or less than another class of data. This problem is difficult to solve; however, solutions such as the oversampling method using synthetic minority oversampling technique (SMOTE) or conditional generative adversarial network (cGAN) have been suggested recently to solve this problem. In the case of SMOTE and their variations, it is possible to generate biased artificial data because it does not consider the entire data in the minority class. To overcome this problem, an oversampling method using cGAN has been proposed. However, such a method does not consider the majority class that affects the classification boundary. In particular, if there is an outlier in the majority class, the classification boundary may be biased. This paper presents an oversampling method using outlier detectable generative adversarial network (OD-GAN) to solve this problem. We use a discriminator, which is used only for training purposes in cGAN, as an outlier detector to quantify the difference between the distributions of the majority and minority classes. The discriminator can detect and remove outliers. This prevents the distortion of the classification boundary caused by outliers. The generator imitates the distribution of the minority class and generates artificial data to balance the dataset. We experiment with various datasets, oversampling techniques, and classifiers. The empirical results show that the performance of OD-GAN is better than those of other oversampling methods for imbalanced datasets with outliers.

One important and challenging issue in handwritten character recognition is the discrimination of visually similar characters. In this paper, we propose a character recognition method for distinguishing similar characters by augmenting commonly used image feature with gradient features from potentially discriminative image regions. The discriminative regions of similar characters sets are automatically detected by analysing the weight vectors of the sparsity promoting logistic fused Lasso method. The histogram of oriented gradients is adopted to compactly represent the gradient features. Additionally, the locality preserving projection method is employed to alleviate the high dimensional nature of the resulting feature vectors. Experimental results on handwritten Lanna Dhamma and Thai characters datasets demonstrate the capability of the proposed method in discriminating visually similar characters. The method also outperforms existing character recognition methods by considerable margins. It has a great potential for character recognition of other alphabets.

A capacitive electrocardiogram (cECG) signal is considered a promising alternative to a conventional contact electrocardiogram (ECG) signal because the cECG signal can serve the same purpose as the contact ECG signal but can be measured during daily life without causing a subject to feel uncomfortable. However, the cECG signal has a limitation in that detection of QRS complexes, which is a fundamental step to analyze heart condition, is not easy. That is because the cECG signal is sensitive to noise, especially motion noise. This paper proposes a method to detect QRS complexes in cECG signals degraded by motion noise. The proposed method is based on fully convolutional networks (FCNs) and mainly consists of three parts: the generation of ground-truth data, the FCN model, and postprocessing. A labeling process for generating the ground-truth data is proposed. Then, an FCN model that is suitable for cECG signals is proposed. The proposed FCN model uses filters of a large size to achieve a large receptive field, unlike the common FCN models used in image processing. The receptive field is sufficiently large to involve information about adjacent QRS complexes, such as the time interval between the QRS complexes and its variability. By considering the information, the proposed FCN model can reliably classify QRS complexes even in cECG signals degraded by motion noise. Additionally, postprocessing, which consists of an accumulation step and a non-maximum suppression step, is proposed to complement the proposed FCN model. In experiments with real data, the proposed method showed an average sensitivity of 96.94%, positive predictive value of 99.13%, and F1 score of 98.02%. These results demonstrate that the proposed method overcomes the limitation of a cECG signal and helps the cECG signal be widely utilized for medical or healthcare applications.

In this paper, a deep convolutional neural network (CNN) is proposed for accurate segmentation of retinal blood vessels. This method plays a significant role in observing many eye diseases. A strided-CNN model is proposed for accurate segmentation of retinal vessels, especially the tiny vessels. The model is a fully convolutional model consisting of an encoder part and a decoder part where the pooling layers are replaced with strided convolutional layers. The strided convolutional layer approach was chosen over the pooling layers approach as the former can be trained. The morphological mappings along with the Principal Component Analysis (PCA)- based pre-processing steps are used to generate contrast images for training dataset. Skip connections are implemented to concatenate features from the encoder part and the decoder part to enhance the vessels segmentation especially the tiny vessels and to make the vessel’s edges sharper. We used a class balancing loss function to train and optimize the proposed model to improve vessel image quality. The impact of the proposed segmentation method is evaluated on four databases namely DRIVE, STARE, CHASE-DB1 and HRF. Overall model performance, particularly with respect to tiny vessels, is primarily influenced by sensitivity and accuracy metrics. We demonstrate that our model outperforms other models with a sensitivity of 0.87, 0.808, 0.886 and 0.829 on DRIVE, STARE, CHASE_DB1 and HRF respectively, along with respective accuracies of 0.956, 0.954, 0.976 and 0.962.

There is no reliable guidance available in literature so far for the selection of a suitable technique for denoising Magnetic Resonance (MR) images. The performance of edge-preserving denoising schemes like Nonlocal Means, Bilateral, Total Variation, Anisotropic Diffusion, Kuwahara, wavelet denoising, Linear Minimum Mean Square Error, Smallest Univalue Segment Assimilating Nucleus and Beltrami filters on MR images are evaluated and compared in this paper. Performance evaluation is done on real-time MR Images, Shepp–Logan Phantom images and simulated MR images. Image Quality Analysis indices used for the evaluation are Structural Similarity Index Metric, Noise Quality Measure, Peak Signal to Noise Ratio, Edge Preservation Index, MetricQ, Anisotropic Quality Index, Blind Reference Image Quality Evaluator and computational time. It has been observed that the performance of each filter is completely different on Shepp–Logan, simulated MR and real-time MR images. It is critically sensitive to the strength of noise also. No filter which can offer good performance equally on Phantom, simulated MR image and real-time MR images, is available in the literature. Values of the objective indices are not in concordance with subjective quality ratings. Filter designs optimized on Phantom or simulated MR using maximum PSNR between denoised and ground-truth images as an objective function (minimum error sense in general) do not perform well on real-time MRI.

Shapelet is a time series subsequence that can best represent the time series of one class. Shapelet can improve the accuracy and efficiency of classification, as well as the interpretability of classification results. Although shapelet has good classification performance, how to efficiently find the optimal shapelet is still an important challenge due to the large number of shapelet candidates contained in a time series. In this paper, a new shapelet discovery method, referred to as Pruning Shapelets with Key Points (PSKP), is proposed. PSKP first finds the key points in time series according to the standard deviation of each time tick of time series, and then extracts shapelet candidates with these key points. Finally, PSKP classifies the time series through a decision tree constructed based on the optimal shapelet. We make experiments on various data sets and evaluate the performance of the proposed method with compared candidates. The experimental results demonstrate that the proposed method is feasible and effective.

Developing an automatic license plate recognition system that can cope with multiple factors is challenging and interesting in the current scenario. In this paper, we introduce a new concept called partial character reconstruction to segment characters of license plates to enhance the performance of license plate recognition systems. Partial character reconstruction is proposed based on the characteristics of stroke width in the Laplacian and gradient domain in a novel way. This results in character components with incomplete shapes. The angular information of character components determined by PCA and the major axis are then studied by considering regular spacing between characters and aspect ratios of character components in a new way for segmenting characters. Next, the same stroke width properties are used for reconstructing the complete shape of each character in the gray domain rather than in the gradient domain, which helps in improving the recognition rate. Experimental results on benchmark license plate databases, namely, MIMOS, Medialab, UCSD data, Uninsbria data Challenged data, as well as video databases, namely, ICDAR 2015, YVT video, and natural scene data, namely, ICDAR 2013, ICDAR 2015, SVT, MSRA, show that the proposed technique is effective and useful.