Search

Paper dating analysis is an important research direction of document inspection, which is widely used in the inspection and identification of cultural relics and ancient books. This article focuses on the analysis of paper dating in ancient books, and proposes a non-destructive inspection and analysis method based on paper fiber texture images. Aiming at the global stacking morphology of paper fibers and the local features of specific types of fiber morphology, we propose a neural network-based hybrid texture feature extraction and representation method: On the one hand, we use convolutional networks to obtain global features of fiber texture; at the same time, we designed a method of extracting and representing local fiber morphological features based on the attention mechanism. By mixing the above two types of features, we realize the extraction and representation of paper fiber features. Furthermore, we use the GRU(Gate Recurrent Unit) model to establish a paper dating time series model and design a new loss function. In order to verify the method, this paper selects 36 domestic books published from 1950 to 2000, and uses the document checker VSC 6000 to collect paper texture images as a dataset, and verifies the effectiveness of the proposed method on this dataset. Experiments prove that the method proposed in this paper has achieved ideal results in paper dating analysis.

In the last several years, we published several papers to address the problem of Face Identification. The techniques employed in those articles were implemented in transform domains. The Discrete Cosine (DCT) and the Discrete Wavelet (DWT) Transforms were utilized, either combined or individually, to extract features which form the final model for each participant in a given dataset. In this paper, we highlight significant parts of our previous works in order to give a fair comparison among all approaches. The results included here are for the following datasets: ORL, YALE, FERET, FEI, Georgia Tech, and Cropped AR. Features are DWT, DCT, energy-based selected DCT-DWT, and combined DCT-DWT coefficients while the classifier is Euclidean distance, either squared or with power of one.

Even current OCR technology has made significant progress with the help of deep learning, rapid identification algorithms for specific information in documents still need elaborate design. In this paper, a lightweight operating framework based on deep neural network is established to deal with the extraction of date and time information from paper worksheets used in transformer station. In the framework, YOLOv4-Tiny is employed to detect the locations of the dates and the time in document images which are mixed with printing and handwritten fonts. Since the handwritten dates in pictures are difficult to segment, a novel lightweight CNN named MobileDateNet is proposed that can recognize the date serially without segmentation. The experimental results show that the identification accuracy of the whole framework reaches 92.1% at the average running speed of 36.5 FPS on an NVIDIA 3090 GPU, ensuring the recognition accuracy and the operating efficiency.

In this paper, we propose a data augmentation framework for Optical Character Recognition (OCR). The proposed framework is able to synthesize new viewing angles and illumination scenarios, effectively enriching any available OCR dataset. Its modular structure allows to be modified to match individual user requirements. The framework enables to comfortably scale the enlargement factor of the available dataset. Furthermore, the proposed method is not restricted to single frame OCR but can also be applied to video OCR. We demonstrate the performance of our framework by augmenting a 15% subset of the common Brno Mobile OCR dataset. Our proposed framework is capable of leveraging the performance of OCR applications especially for small datasets. Applying the proposed method, improvements of up to 2.79 percentage points in terms of Character Error Rate (CER), and up to 7.88 percentage points in terms of Word Error Rate (WER) are achieved on the subset. Especially the recognition of challenging text lines can be improved. The CER may be decreased by up to 14.92 percentage points and the WER by up to 18.19 percentage points for this class. Moreover, we are able to achieve smaller error rates when training on the 15% subset augmented with the proposed method than on the original nonaugmented full dataset.

People re-identification is a fundamental stage in video surveillance systems. This stage is used for determining the labels of the people’s images usually based on their visual similarities. The appearance salience caused by wearing the clothes and carrying objects, with specific colors can be used for improving the performance of re-identification systems. To address the mentioned point, in this paper, we propose a color-based re-ranking process. In the proposed approach first, the input images are semantically segmented as: person’s body, possible carried objects, and background. Then the most frequent colors of person’s body and possible carried objects are used to re-order the ranked results obtained from two commonly used re-identification approaches i.e., Gaussian of Gaussian (GOG) and Hierarchical Gaussian Descriptors (HGDs). Experimental results on various people re-identification datasets show that our proposed re-ranking approach can improve the performance of re-identification systems.

This paper is review on image fusion for different technique i.e. Non-subsampled Contourlet Transform (NSCT) and Neural Network (NN). NSCT is the pyramid structure and gives the multiscale feature of the image. NSCT used to directional filtering structure and used to filter bank. Neural network is a progression of calculations that tries to perceive fundamental connections in a bunch of information through an interaction that impersonates the manner in which the human mind works. The implemented algorithm will applied to MRI image and with the help of NN to calculated accuracy. The previous algorithm was work of different technique i.e. principal component analysis, discrete wavelet transorm and curvelet transform. The implemented algorithm will implement MATLAB software and calculate mean square error and peak signal to noise ratio.

The aim of this work is to classify the dementia stage using microstructural white matter (WM) diffusion kurtosis indices. We develop various class imbalance machine learning methods in classifying the cognitive normals (CN), Alzheimer's disease (AD) and mild cognitive impairment (MCI) individuals. Diffusion-weighted images of 155 subjects were collected with 44 CN, 88 MCI and 23 AD individuals aged between 60 to 96 years. We first perform skull striping using FSL tool along with head motion and artifact correction. We calculate the mean diffusivity, fractional anisotropy from various brain regions of WM indices from diffusion kurtosis imaging (DKI) using tract based spatial statistics. Then we estimate the kurtosis fractional anisotropy of various WM regions. After estimation of all the regional KFA indices, we trained these indices using various class imbalance machine learning classification algorithms, as the sample size is different in all the three groups. We found that balanced random forest classifier was the best classifier with an accuracy of 74% when compared to other methods for 5-fold cross-validation. We conclude that class imbalance machine learning methods are potential in classifying cognitive normals, AD and cognitive impairment individuals.

Image classification needs to consider image degradations in practice because an image classification network trained with clean images works poorly for degraded images while digital images usually include some degradations such as JPEG compression. To tackle this problem, a common approach is training classification networks on degraded images with various levels of degradation, e.g. various quality factors for JPEG compression. However, the classification networks do not usually have enough accuracy for clean images because the classification networks have been averagely trained for various levels of degradation. This paper aims to construct a classification network of degraded images without sacrificing the performance of clean images. This paper proposes a network to learn the classification of degraded images and degradation levels of degraded images as multi-task learning. Learning the classification of degraded images is the main task of multi-task learning. On the other hand, learning degradation levels is a sub-task of multi-task learning and reinforces the classification ability. In the proposed network, a feature extractor of the classification is trained with image features acquired from a classification network of clean images by using consistency regularization with the cosine similarity loss. The experimental results using different types of degradations, including JPEG compression, Gaussian blur, Gaussian noise, and salt-and-pepper noise, show that our proposed network has enough ability to classify degraded images without sacrificing the performance for clean images.

© 2021 IEEE.Coronavirus usually transmits from the animal to the human, but now, the virus transmission is between persons. Therefore, scientists and researchers are trying to develop several types of machine learning methods to defend against COVID-19. Medical images play a significant role in this time due to they can be used to recognize COVID-19 accurately. However, in this paper, we used X-Ray images, the images undergone to sharpening techniques to increase the results further. The texture techniques named local binary pattern (LBP) have been used in order to extract features. The features obtained were applied to the support vector machine (SVM). The results we achieved were 100% for all performance measurements. Our results were conspicuously superior compared to the state-of-the-art papers published.

In this paper three different types of architectures have been trained and tested to determine which one of them is more efficient in detecting exudates from retinal images. The database that was used is SUSTech_SYSU from which 100 images were extracted for image testing and another 100 images that were used in obtaining the training dataset. The three architectures that were studied in this paper are SqueezeNet, ResNet50 and GoogLeNet. The results showed that from the three architectures the most efficient in exudates detection was ResNet50 with an average accuracy of 0.928, followed by SqueezeNet with 0.910 and finally, GoogLeNet for which the average accuracy was 0.907.

Video surveillance is an important method to obtain the dynamic changes of green macroalgae along the coast. The paper proposes a coastal green macroalgae extraction method based on the SLIC superpixel segmentation, CNN and SVM to realize the automated recognition of green macroalgae from lots of high-resolution RGB video data collected by unmanned aerial vehicle (UAV) and handheld devices. Firstly, SLIC algorithm is used to generate the multi-scale patches on the original high-resolution image. Then, three classification CNN is used to divide the multi-scale patches into three types: green macroalgae, background and mixing. Finally, SVM algorithm is used to extract the green macroalgae to improve the accuracy at the pixel level in the mixed patches. In order to evaluate the performance of the proposed method, experiments are conducted on our coastal green macroalgae image dataset. Compared with the method of RGB vegetation indices (such as ExR, RGBVI, NGBDI), the overall accuracy (OA), F1 score, and Kappa of the green macroalgae extraction with the method proposed in this paper are up to 95.23%, 0.9612, 0.9436, respectively. The results show that our method is significantly better than that of RGB vegetation indices since it effectively reduces the influence of sea waves and light on the recognition results. The automated extraction method for coastal green macroalgae proposed in this paper can provide a reference for the automatic monitoring of coastal green macroalgae with high precision.

Pneumonia is one of the most common infectious diseases in clinic. X-ray chest is an important basis for early diagnosis of pneumonia. With the development of computer vision technology, using convolutional neural network to train pneumonia image classification model has been gradually applied to the process of medical clinical diagnosis. However, there are many problems in the process of using convolutional neural network to train pneumonia image classification model, such as too long model training time, over fitting and low accuracy due to too small training dataset. To solve these problems, this paper proposes an optimization method of pneumonia image classification model based on transfer learning and feature fusion, which is called Transfer Fusion. The Transfer Fusion optimization method will transplant the trained source model parameters to the target model, and add a specific feature fusion classification layer, so as to significantly shorten the training time of the new model, improve the accuracy and prevent over fitting. In this paper, Transfer Fusion optimization method is applied to three common convolutional neural network models: Google InceptionNetV3, MobileNetV2 and ResNet50. Through a large number of experiments, the performance of the three models has been significantly improved and improved.

Diabetes mellitus being a frightening illness, caused due to higher glucose levels in the blood. The suitable usage of machine learning (ML) algorithms assist in the early diagnosis of diabetic mellitus. The primary aim of this paper is to utilize the simple metaheuristic technique, the crow-search optimization algorithm for better diabetes prediction. In this regard, the paper employs the Pima Indians Diabetes dataset. The crow search algorithm used in this work is for transforming the non-linear data input of the preprocessed PID dataset to substantial linear ones. Later, the classification is done using ML techniques namely K-Nearest Neighbors, and Random-Forest algorithms. Thus the work utilizes the crow-search optimization algorithm as a transformation-based technique to improve the classification results of the diabetic problem. Finally, the performance of the above ML techniques with and without transformation using the crow-search algorithm is analyzed and assessed by employing benchmark performance measures namely accuracy, F1 Score, and precision. And as a final point, the obtained results have been validated using the Matthews Correlation Coefficient. To this point, the proposed transformation technique with the Random Forest algorithm for the diabetic classification problem provides superior performance of 94.27% accuracy over others.

The description of gene intonation can be exploited in the determination of acute disorder like carcinoma. The picking of biological genes is substantial as well as vital for cancer observation. This paper offers an outline suggesting two phases of gene selection for colon data categorization. Colorectal tumor positions third among extensive cancer and foremost reason of cancer demise globally. There is a necessity to realize exact process to recognize the tissues and to improve the diagnostic method for discovering medicines. Freely accessible binary class colon data that embraces 2,000 genes and 62 trials is used in this research. The two-phase feature selection includes Analysis of Variance test as first phase and in second phase, the Hilbert transform and Firefly algorithm are applied to get pertinent genes. Then finally, the obtained colon data are classified as normal and tumor classes using Linear Regression, Logistic Regression, Naive Bayes and Harmonic search classifiers. The results indicate that the Harmonic search classifier provides a greater accuracy of 95.06% for Hilbert transformed colon tumor features and an accuracy of 90.37% for Hilbert transformed colon normal features. For Firefly clusters, the naive bayes classifier attains a higher accuracy of 93.36% and 91.93% respectively for colon tumor and normal data. It is deduced that the performance of two-phase gene selection method with Harmonic search and Naive bayes classifier has accomplished a better accuracy in contrast with the other classifiers employed in this paper.

Residual convolutional networks proved superiority in the image recognition field in addition to many other pattern recognitions related problems. A special benefit of residual convolutional networks has been proven in the field of handwriting recognition, especially in handprinted characters and online signature recognition. Two published research papers addressed the two problems with highly accurate results that compete with the state-of-the-art techniques. Both problems have two well-known challenges in object classification: intra-class variability and inter-class similarity. In the character recognition problem, different users produce different shapes for the same character in addition to the user himself, who produces different shapes for the same character resulting in high intra-class variability. Inter-class similarity exists due to the similarity between different characters. In handwritten signature recognition, intra- class variability is introduced by the same person where no two signatures of the same person may coincide. On the other hand, persons may have similar signatures resulting in inter-class similarity. Solving the two pattern recognition problems using deep learning techniques requires the selection of the best technique that can model the handwriting process, but at the same time, it requires a large training data. In the character recognition case, large training samples are available from many datasets resources. A recent data set representing this challenge is the EMINST dataset. An optimized residual architecture has been introduced to give an excellent solution for this problem in one of the comparison papers. In the case of signature recognition, a different solution based on a residual network has been introduced in a second paper because practically training samples that can be collected from new users are very rare. In this paper, the two techniques are compared, and conclusions that may generalize well in other problem domains are stated.

Epilepsy is a central nervous system disease that affects loss of consciousness and convulsions. The testing of electrical activity in the brain is identified as EEG. Ability to predict the onset of epileptic seizures before their commence seems to be very useful for medication-assisted seizure protection. Electroencephalogram (EEG) signals are now used to predict epileptic seizures using machine learning techniques and statistical analysis. Pre- processing of EEG signals for dynamic range and extra functionality, but in the other hand, extraction are two real issues that have a negative effect on the system’s run time and correctness. Therefore, this paper proposes a model that provides a reliable method for both pre-processing and classification. It is implemented using FIR filter for preprocessing and machine learning classification methods. Among the variety of machine learning algorithms, this paper uses RF algorithm and BPNN for classification. After the classification, the results were compared, and it is revealed that Random Forest classifier gives 95% accuracy which is more efficient than Neural network via back propagation.

The electrical action of the heart is recorded by an electrocardiogram signal (ECG). It records the electrical signal in order to check for various heart ailments and disorders. The precise computer analysis of ECG signals is difficult, as being a low-frequency signal, it is exceedingly vulnerable to high-frequency noise and numerous other losses. ECG filtering is therefore a major pre-processing step that attenuates the signal noise and converts it into a signal better suited for analysis. Researchers over a while have given various techniques to correctly detect and filter out the noises. In this paper, a number of state-of-the-art techniques for ECG signal pre-processing using FPGA on MT-BIH Arrhythmia and online datasets are used. There are a number of digital filters categories likewise FIR, IIR, and adaptive filtering techniques are present. The performance of these techniques are analysed on three parameters: Power Consumption, Signal-to-Noise Ratio, and Accuracy. It has been analysed that, the median filter attains the best SNR value by employing Neural Network-based ECG on FPGA board. From the analysis of the papers, we find that the median filter provides the best SNR value, Neural Network-based ECG Anomaly Detection on FPGA technique provides the best accuracy and the combination of FIR LPF & HPF using Kaiser and Bartlett windowing technique provides the lowest power consumption.

Human activity recognition is a challenging problem of classifying human activities based on video streams or accelerometer data. Most activity recognition methods use depth enabled data or multi-modal data to precisely predict movements. As majority of the real-world applications of activity recognition use RGB cameras, this paper aims at using RGB-only videos to classify actions. Motivated by recent advances of 1-Dimensional Convolutional Neural Networks (CNNs) to classify sequential data, this paper investigates the application of a deep 1D-CNN model for activity recognition as compared to a Bi-Directional Long Short-Term Memory (BLSTM) method. Instead of using raw RGB frames as input to the network, we utilize OpenPose API to detect 2D skeleton key points in video frames, which are then used to recognize human actions. This paper also addresses the challenges of training deep models with limited labeled data by using data augmentation and dynamic frame dropout techniques to increase the efficiency of the model and avoid overfitting. We verify the performance of our model on three popular and challenging benchmarks: UTD-MHAD, KTH and UCF-Sports. Extensive experiments demonstrate that our 1D-CNN model outperforms the BLSTM model and all the state-of-the-art methods consistently.

Micro Expressions are those involuntary muscular movements of the facial muscles produced in response to a stimulus. They are short-lived expressions that last for anywhere between 0.04 to 0.2 seconds and are extremely subtle in their amplitude. Given their fleeting and elusive nature, it becomes almost impossible to detect these expressions through the naked eye. Recent developments in Deep Learning models have shown great success in efficiently identifying and analyzing Micro Expressions. In this paper, various models have been implemented on the SAMM dataset. The models studied are namely– VGG16, ResNet50, MobileNet, InceptionV3, and Xception. The experimental results have helped us carefully analyze the various metrics related to the models and compare them with each other to ascertain which one outperformed the others and is best suited for real-world applications. The MobileNet model has surpassed all other models in terms of its efficiency with respect to the domain of this paper. It has been able to describe and understand all the information that can be found in the various Micro Expressions.

For humans, text is one of the most significant sources of knowledge. Detection of text in images has been an interesting computer vision topic in both industrial and academic study. There has been a lot of effort made into improving existing methods and establishing new ones. Sliding windows and connected component systems are more classic solutions.In this paper we are discussing some text detection methods.In this paper we are discussing text detection and identification based on convolutional neural network.Script identification is done by integrating local cnn and global cnn.

Corn yield improvement program aims to attain continuous national self-sufficiency. The program needs to be supported by the availability of food resources, including high-quality corn seeds. In corn seed production, grading is one of the factors that affect the quality of corn seeds. The grading process is conducted manually by visual observations of workers. This process tends to be subjective and ineffective. Some corn seed factories use sieve machines to do grading by seed size. In this paper, an imaging-based classification system is proposed to perform corn seeds (BIMA-20 URI Hybrid) grading of two classes, which are categorised as good and bad. Three different methods are studied in the paper. The methods are respectively based on (1) shape, colour, and size features, (2) seed roundness, and (3) deep learning approach. Images data is acquired in a group of five corn kernels. Region-of-interest (ROI) segmentation is performed to select every single seed from the group image. Features values are then extracted from a single seed image and used as a classification parameter. The F 1 score of the proposed classification system, roundness differentiation, and model training performance can be used to show the categorisation capability. The deep learning approach has achieved the best F 1 score among the other proposed techniques. The best F 1 value, 0.983, is obtained at the ResNet-50 implementation. In separated observation, Method 6 (Size and Colour), Method 7 (Size, Shape, and Colour), Roundness, and ResNet-50 are represented as the best model for each group method. These methods reach F 1 scores more than 0.9, except the roundness parameter. The F 1 score of the roundness parameter is found at 0.854. Additional parameters might be required by the method based on the roundness feature for improving its final performance.

Carbonate rock is a kind of very valuable sedimentary rock, and oolites are one of the most easily identifiable particles in the carbonate rock image. Based on the traditional image segmentation algorithm, this paper proposes two extraction strategies for carbonate oolitic components. The first is the extraction technology of carbonate rock oolites based on traditional image segmentation algorithm: first extract the connected domains of the carbonate rock image, then use the K-Means clustering algorithm to analyze the processed image, and then analyze the image Binary processing, and finally extract the contours of the oolites; the second, the carbonate rock oolite extraction technology based on the superpixel segmentation algorithm: first use the SLIC algorithm to segment the acid rock image; secondly, the road extract. The test results show that the two extraction strategies can clearly extract the oolitic components of salt rock. In addition, according to the experimental results in this paper, it can be seen that the oolitic extraction technology based on SLIC superpixel segmentation is slightly better than that based on traditional image segmentation algorithm.

At present, many challenges exist in the application of waste image classification including high image resolution, background noise interference and low computational efficiency, which makes a great impact on the accuracy of the algorithm. Generally, convolutional neural network related strategy is adopted, which can improve accuracy to a certain extent. But it is prone to getting interfered by irrelevant information and has large size of parameters, which makes it difficult to be deployed for those resource-limited applications. Thus, a Vision Transformer is proposed. By using Vision Transformer, it significantly improves the anti-interference ability and reduces parameter scale. Furthermore, distillation strategy is adopted to improve the generalization ability of the model. In addition, this paper replaces multilayer perceptron (MLP) layer in base Vision Transformer with improved Locally-enhanced Feed-Forward (improved LeFF) to improve the convergence speed and accuracy of the model. This paper also proposes Vision combinative self-attention algorithm to reduce the requirements of computational resources especially for image with high resolution. The results show that in 79002 waste image datasets with 198 types, the accuracy of training set and test set can reach 99.98% and 95.03% respectively. Meanwhile, a relatively high accuracy rate could be achieved within less training epoch and less computational resources are required during the operation.

Emotion recognition is a growing area of facial recognition, to detect the basic emotion state of a person and then operate further analysis. For practical applications, high speed and accuracy are required as an efficient and precise system. To this end, the paper proposes an effective emotion recognition system using a representative geometric feature mask for feature extraction and a CNN model for classification. Compared with traditional emotion recognition systems, which usually extract facial key features and then convert them into mathematical information variables by equations, the system implemented in this paper extracts necessary features in facial expression through landmarks, and operates a further extraction by a transformation that converts features into a pure geometric feature mask to represent a simplified human face. Then, the mask that can be used to express the human facial emotion with fewer noise features, is input into a deep learning training CNN (Convolutional Neural Network) model. The improvement of this work is that the system combines pure geometric method to extract facial features with CNN algorithm properties in image processing, where local connectivity and shared parameter properties were fully used in further geometric feature extraction. Finally, the system achieves high accuracy and low time costs with KDEF (Karolinska Directed Emotional Faces) and CK+ (Cohn-Kanade AU-Coded Expression Database).

In this paper, a method of industrial time series data cleaning using generative LSTM model and adaptive confidence interval is proposed. Firstly, the generative LSTM model is used to predict the probability distribution of industrial time series data, and then the confidence boundary of data anomalies is calculated according to the adaptive confidence interval method, so as to judge whether the data is abnormal or not. This paper uses network open data to complete data cleaning experiment, and the precision rate and recall rate of abnormal data detection are above 96%. The comparison with other anomaly detection algorithms also proves the effectiveness of the proposed algorithm in industrial time series data cleaning.

Deep learning (DL) is a fresh study orientation in the field of machine learning in computer science. It is recommend into machine learning to make it nearer to the customary artificial target intelligence. DL is the inherent law and express level of learning sample data, and the information get in the learning procedure is of mighty help to the explain of data such as words, images and sounds. CNN (Convolutional Neural Network) combines feature extraction with itemize process to train neural network, which has acquire mighty successful in the field of image classification. This paper focuses on the automatic classification of fetal facial ultrasound images. A 19-layer convolution network is proposed and improved. By using data enhancement, adding global mean pooling layer, reducing the number of channels in the full connection layer of the model, and optimizing learning based on parameter transfer learning of fine-tuning training, the automatic classification of fetal facial ultrasound images with limited data volume can be realized. Match with the present solutions, the depth network proposed in this paper can effectively avoid ultrasonic noise interference and learn deep features more effectively. A heavy quantity of specific analysis test have proved its effectiveness.

Convolutional neural network has achieved a lot of success in the field of computer vision in the recent years. With the rapid development of convolutional neural network, most image classification task has achieved significant performance improvement. Although some progress has been made in the research of image classification methods, there are still some deficiencies. For example, many existing methods are difficult to adaptively mine the feature importance within the sample and the feature correlation between sample scales. In order to solve the above shortcomings, this paper mainly studies image classification based on dynamic adaptive learning. In this paper, Multi-Scale Dynamic Convolution (MSDC) is proposed and verified on the standard image classification data set. Our method can be adjusted adaptively according to different scales of input data. The experimental results show that the proposed method exceeds the relevant comparison methods.

Skull fracture detection is to detect and locate the area in the skull images. Because the skull fracture areas are relatively tiny, it is generally difficult to detect. Moreover, the fracture areas are easy to be confused with normal sutures. Hence, skull fracture detection is a small object detection. For small object detection in the medical images, the object detection methods based on deep learning is difficult to achieve satisfactory results. In order to further improve the accuracy of skull fracture detection, this paper proposes a novel skull fracture detection method based on Feature Pyramid Network (FPN) with the attention mechanism. This method uses the criss-cross attention mechanism to highlight the contextual information of the objects in each layer of the feature maps in the feature pyramid, which effectively enhances the target information in the feature pyramid. Because data augmentation can effectively improves the detection accuracy, this paper proposes a new data augmentation strategy to improve the detection robustness of the network. In addition, in order to more accurately reflect the locate effects of candidate boxes, this paper proposes a new Intersection over Union (IoU) calculation method, which is based on the distance between the center points of the predicted boxes and the labeled boxes, adding a scale factor to the original calculation method of IoU. The new method can better reflect the actual situation. We use the skull fracture dataset provided by the hospital to verify the proposed method. The experimental results show that the proposed detection method outperforms the current mainstream object detection methods based on deep learning in detection accuracy and recalling rates.

Epilepsy is a common neurological disease at present. It is caused by abnormal discharge of brain neurons. The clinical manifestations are whole body convulsions, loss of consciousness and stopping breathing. By using electroencephalogram (EEG) to capture changes in brain electrical activity, many studies use EEG signals for epileptic seizure prediction. This paper proposes to predict epileptic seizure EEG signals based on a Gated Recurrent Unit (GRU) prediction model. First, the original data set is passed through a 50HZ notch filter to filter out the noise and artifacts that may be included in the data collection, and then it is sent Enter the GRU prediction model for prediction. This paper uses both LSTM and random forest model algorithms. Experiments show that the GRU model algorithm has the shortest training time and the accuracy rate is 12.48% higher than that of random forest, which proves that the method is feasible.

The SAR image port region semantic segmentation algorithm combined with the deep learning mechanism is of great significance to the port intelligence and information construction. However, there are few researches on the pixel-level semantic segmentation of port area with SAR images. The port spatial pattern monitoring based on remote sensing images still has problems such as complicated monitoring procedures, low robustness, and insufficient datasets. To take advantage of SAR images and explore the potential of convolutional neural networks in the SAR port area images, a dataset containing eight land types of port was established, and a pixel-level segmentation framework for port areas based on SegNet was proposed in this paper. The pixel-level classification of high-resolution SAR port area images based on small datasets was realized. The comparison experiments with SVM and random forest algorithm proved that the algorithm proposed in this paper can achieved an overall accuracy of 91.6% on a SAR image of port area, and better maintain the edge characteristics of different land types.

Due to the multi-view imaging principle of synthetic aperture radar (SAR), its imaging process is not limited by any time or any bad weather. The detection of marine ship for SAR image is a very important application in both military and private applications. A marine ship detection method in SAR image about an improved Faster RCNN-based approach is proposed. First, the deepest semantic features of SAR image extracted by Faster RCNN contain less target information and small targets may be ignored, so that we fuse the deepest feature and shallower features in the feature extraction network; Secondly, because the redundant information in the features will produce false alarms, we embed the Convolutional Block Attention Module (CBAM) in the feature extraction network to extract more effective features; Finally, we use bilinear interpolation to obtain floating-point coordinates to optimize RoI pooling which uses rounding quantization, so the error mapping caused by quantization will be reduced. The algorithm of this paper is tested on the SSDD public dataset, and the AP is increased from 0.79 to 0.89. The results demonstrate that the algorithm introduced in this paper has a very significant performance improvement for detecting ship-like targets from SAR images.

Recently, 3D palmprint recognition systems have started to gain the attention of researchers compared to their 2D counterpart. The key task in the design of a biometric identification system is the choice of the method of extracting discriminating information from 2D/3D images. Given the enormous success of deep learning approaches in the field of feature extraction and classification, we propose in this paper the use of Transfer learning networks in order to build a palmprint based person identification system. Also, this paper aims to show the efficiency of feature selection to improve system performance. To do this, a Relief method was used to perform the feature selection task. Experiments on a 2D/3D palmprint database with 8000 samples show that the proposed scheme can significantly improve the efficiency of palmprint identification.

The digital halftoning is widely adopted in printing applications to obtain a high-quality printing output with optimal ink usage. The technique comprises of transforming a continuous tone grayscale image into its approximate binary image. On the other hand, image quality assessment deals with estimating the perceptual quality of the degraded image. The paper deals with exploitation of ordered dithering technique of the halftoning, to improve the performance of the existing image quality indices. The widely used full reference image quality metrics such as Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are considered for this paper. The main objective aims to improvise the correlation coefficient between the widely used IQM and Human Visual System (HVS). Moreover, the computation demand of the existing indices is also reduced significantly. The well-known LIVE image quality database is used to validate the performance of the proposed approach. To perform correlation analysis, Pearson, Spearman and Kendall correlation coefficients are used. From experimental analysis, it can be validated that the proposed IQM performs with improved accuracy and less computational demand.

Medicine, and particularly radiology, is an area where vision systems bring significant benefits, which results in more accurate diagnoses, predictions, and treatment plans. This paper proposes a U-Net training model and simplifies the deep learning based framework for tumour detection and semantic segmentation that shapes frames for computer-aided diagnoses (CADx) and computer-aided detection (CADe) applications. The U-Net was used to segment glioma – tumour area. A training technique is proposed and convolutional neural network models capable of being trained on a dataset of fused full-size CT and PET scans – 512x512x1 – is addressed. The implemented algorithms were executed in a cloud environment, where storage was decoupled from compute (CPU and GPU). The resultant models’ performance was assessed utilizing multiple metrics. Trained models fully automatically detect tumours in a given dataset. The best scored coefficient (Dice Co-Eff) for a model is 0.8750. The paper which follows is a detailed reference with training parameters for U-Net for efficient brain tumour detection on CT and PET scans where the dataset is limited –- in this case, consisting of merely 20 patients.

Electroencephalography (EEG) signals are normally susceptible to various artifacts and noises from different sources. In this paper, firstly the existence of artifacts will be identified on the recorded EEG signals and then the origin of the detected artifact will be determined among 7 different sources. Due to the nature of EEG signals, almost no specialist can determine artifact source through eye inspection. This paper introduces the utilization of 1-D Convolutional Neural Network (CNN) in multi-class EEG artifact classification. Proposed CNN models were kept as simple as possible to have the best operation time but in the meantime, models were selected adequately deep to extract appropriate artifact features from applied EEG signals. Obtained results prove that proposed architectures are able to classify artifacts with high accuracy.

In visual understandings, images taken from different cameras usually have different resolutions, illumination, poses, and background views that lead to domain shift. Besides labeling these data is an expensive operation. These problems lead to the need for unsupervised domain adaptation (UDA), in which training and testing data are not drawn from the same distribution, and labels are not available in the target domain. This paper presents an improvement for unsupervised domain adaptation in transfer learning using a unified discriminant and distribution alignment (UDDA). The existing method of UDA only utilized unsupervised PCA as the dimensionality reduction process before being added to the joint objective function consisting of distribution discrepancy minimization and regularization. However, the label in the source domain has been utilized by some works (i.e., joint geometrical and statistical (JGSA)) to use the supervised method LDA and show good improvement. Nevertheless, LDA has some drawbacks that is sensitive to noise and outlier in square operations and only take part in global information. The contribution of this paper is to add local discriminant information to the proposed UDA model by adopting locality-sensitive discriminant analysis (LSDA) that strengthens the between-class and within-class discriminant in the source domain. In this method, within-class and between-class graphs will be computed, and summed up with within-class and between-class scatter matrices before being embedded into the joint domain adaptation framework. Comparing the proposed method with the state-of-the-art techniques in objects and digital datasets, it shows our UDDA improved the average accuracy by 3.43% and 7.13% compare with the second and third highest results respectively.

This paper presents the deep learning model by Convolution Neural Network (CNN) in training the crop disease classifier via image classification. A camera will be equipped and applied in artificial intelligent drone to operate as a crop monitoring system used for agriculture. Agriculture productivity is a key component of country economy. Crop diseases can lead to a drop in the quality and quantity of agricultural products. Famers are facing problems to detect the crop diseases accurately in huge region of crops. Therefore, CNN based method for crop disease detection is proposed. Dataset contains of 16,257 color images which has a total of categories have been fed into the model, out of which 10 categories are of diseased crop leaves. The CNN model contains 7 convolution layers with the number of filters 32, 64, two layers with 128 filters, three layers with 256 filters and filter size $3\times 3$ is the proposed approach to perform crop disease classification, with the best testing accuracy of 99.02%. The crops are classified correctly using the suggested CNN design. The suggested CNN design is validated and evaluated which achieves accuracy of 99.02%, 0.98% error, 99% recall, 99% precision and 0.99 score of F-measure. In this paper, achievement of the proposed CNN model is reaching a promising result and simulated successfully in classifying the crop disease.

Generative Adversarial Network (GAN) is an algorithmic architecture containing two neural networks, placed against each other to generate new synthetic images and it has been used successfully in image segmentation. The paper analyzes different GAN implementations for segmentation of images acquired by aerial robots in a real context of a rural zone in Romania. To improve the segmentation performance, a new GAN network is proposed by adding a new layer. Data augmentation was done by the following techniques: mirroring, rotation, scaling, gray scaling, blurring, sharpening, etc. Five classes of region of interest are considered: floods, vegetations, buildings, roads, and dry land. GAN implementations were tested on CPU, GPU, and TPU, on individual computing devices and in the cloud. A new layer was added. The performances were analyzed in terms of learning time, operating time, and statistical indicators. The batch size was generally low: batches of 1, 4 or 16 images were used in this paper. The results confirm that the use of batch achieves the best training and generalization performance in terms of computational cost, for a wide range of experiments.

The purpose of this paper is to describe a novel fusion algorithm for multi-modal biometric identification. In this paper we describe the fusion of fingerprints and voice. This combination of biometrics is rarely used in verification systems although this biometric pair is simple to use and not too invasive. A framework for the combination of several data fusion algorithms is described. In this paper we use only two types of data fusion techniques, namely weighted sum and fuzzy system. Two independent identity decisions can be thus obtained, and from them two beliefs that the identity is verified can be derived. The two beliefs are combined using Dempster-Shafer's approach to obtain the final decision. The results are reported by ROC curves.

This paper presents a modified image and video analysis model designed to detect nonhomogeneity and anomalies in the marine environment. Unlike known alternatives, the model can detect natural objects after being trained on scarce data and having access to limited computational resources, in anomalous visual monitoring data collected directly in the marine environment. Accuracy in terms of type I and II errors has been improved by implementing additional preprocessing using intelligent anomaly detection and elimination methods. Requirements for the diversity of training sets have been lowered significantly, whilst the use of invariant metrics with adaptive thresholds improves detection adequacy in cases of affine distortions of scaling and rotation of monitored objects. The paper further compares the results of numerical experiments run on real-world data of in-situ monitoring of marine objects.

The paper describes an approach to constructing image models using discrete wavelet transforms. The models under consideration reflect the importance of pixels in terms of their contribution to the total energy. In this case, each pixel is associated with a certain numeric value – weight. To obtain the weights, the detailed coefficients of the wavelet transform are analyzed, which reflect the magnitude of the brightness change in the corresponding pixels at different scale levels. It is assumed that the more the brightness changes in a pixel, the more this pixel differs from other pixels and the more important it is for image analysis. The values of the weights depend on the type of the wavelet transform. In this paper, we consider orthogonal Daubechi 2, 4, 6, 8 wavelet transforms and biorthogonal Villasenor-Belzer-Liao 6/10, 5/3, 2/6, 9/3 wavelet transforms. The application of weight models based on the specified wavelet transformations for the selection of texture features of images is described.

High mortality rate among women in India is mainly due to breast cancer. As mammography is a standard imaging tool used in screening the breast, it sometimes tends to miss certain abnormalities that lie beneath the breast regions. Thermography being a temperature-based functional imaging modality has enough potential to identify the abnormal changes in the breast through asymmetries at an early stage. This paper considers breast thermograms of patients that are normal and abnormal wherein pre-processing is applied to extract the ROI. The pre-processing includes breast region segmentation using horizontal and vertical projection profile approach along with the Otsu method. The features are then extracted from this ROI for which they are also fed as input to the non-linear classifiers for classification of normal and abnormal. The outcome of the paper results in a good accuracy rate that can be efficiently utilized as a Computer-Aided Detection tool for a second opinion in the breast cancer diagnosis.

This paper proposes a framework that will detect different types of traffic signs from images for Traffic sign recognition and detection in advanced driver support system. In this paper classical LeNet5 network model is used with GTSRB dataset for evaluation purpose. Shape and color are the two main factors that distinguish road signs from other objects. Firstly, preprocessing of an image to highlight important information. Secondly, the training data are scaled, normalized, expanded, and augmented. Finally, the validation and testing phase in which traffic signs are classified and accuracy is calculated. It is concluded from the results that our system has achieved greater accuracy due to the architectural optimization of the neural network for detecting traffic signs. Significantly better performance was achieved with partially blurred images. Traffic sign are classified in 43 classes. For training 22271 images are used in which 6960 images used for testing and 5568 images are used in the evaluation purpose. With new design model 98.75% accuracy is obtained.

In this paper, a brain tumor segmentation method has been proposed. Image segmentation is required for the detection of brain tumors. For this purpose, brain images are divided into two distinct areas. This is one of the main but also the most complex element of the tumor identification process. A brain tumor is a result of mass of tissue that grows, it is the most important cause of the increased mortality rate among children as well as adults. This paper proposes a brain tumor segmentation method by using a combination of Adaptive Histogram Equalization and U-Net architecture. The performance of U-Net architecture and U-Net Architecture with Adaptive Histogram Equalization Technique are contrast by using PSNR (Peak Signal to Noise Ratio), MSE (Mean Squared Error), Quality and Complexity parameters. Overlap-tile strategy is used to smooth the segmentation of large images. Many MR images were collected, and the observations were carried out for statistical analysis using the Adaptive Histogram Equalization and U-Net architecture.

Person classification based on gait postures has attracted increasing attention in the biometric field. The human gait posture has robust discriminative features and it can classify a person even at a distance. The gait posture features also strongly discriminate in low resolution. Unfortunately, the view-invariant characteristic degrades the performance. To address this issue, we propose a comparative study based on model-free techniques. The latest techniques based on view-invariant model-free approaches such as Normal Distance Map (NDM), CNN with Angle Centre Loss (ACL) and Attentive Spatial-Temporal network are used in the study. The paper surveys extensive existing popular techniques towards model-free approaches for person classification. This paper discusses the performance analysis and gait features that help to classify a person from a different view angle. Comprehensive experiments on CASIA-B gait dataset firmly demonstrate efficacy for cross-view person classification.

The respiratory sound classification paper focuses on analyzing sounds collected through the stethoscope at various chest positions to understand critical concerns of human breathing. Most doctors used a stethoscope for auscultation of lung sounds, along with heart sound, as a basic checkup and utilized later in predicting chest pathology. Using deep learning-based sound classification can build intelligent digital stethoscopes or understand overall health conditions later used along with other Lung organ analyses. The solution enables affordable and practical tools for remote parts of India, where doctors or specialists are not keen to work. Such solutions can be used as first analysis before further investigation by traveling to far-off specialist hospitals. The lung sound was prepared and hosted by a team from the University of Coimbra, Portugal; the University de Aveiro, Portugal; and the Aristotle University of Thessaloniki, Greece and the collaboration at Kaggle website lung sounds of patients who are Normal lung functioning and with some anomaly collected. The collected lung sounds are labeled as Normal, seven classes - Abnormal. Overall, eight levels are considered to classify using a deep learning model. The paper proposes a multipath convolution neural network (CNN) model for analyzing sound after exploring all approaches, like Recurrent Neural network (RNN), Machine learning models like SVM, and other CNN models as per paper [3]. Lung sound is processed using signal processing methods and converts wav files to Mel-frequency cepstral coefficients (MFCCs), Mel spectrogram, and chromatogram from a waveform or power spectrogram. Instead of using one feature among three, the paper built a Multipath CNN model to feed all three components to model and concatenate feature maps to classify the classes. The model was evaluated with various parameters like accuracy and F1 score. The proposed model gives 96.17% for the dataset used. The result implies data, and a multipath approach helps to fuse features extracted from different images and allows models to classify well.

Facial expression recognition has long been a challenging task in the computer field. Most of the existing facial expression recognition methods are the extraction of facial features, but these methods emphasize that the neural network should pay attention to some key features of the expression, such as eyes, nose, and mouth. These methods ignore some minor features that also affect expression recognition, such as facial texture and other features. This paper proposes an efficient Shuffle Attention Supplement (SAS)-Net attention module to solve some minor facial feature extraction problems. SAS-Net mainly combines the feature map before entering the SAS attention module and the attention feature map formed after aggregation of the submodules in the SAS attention module. This method makes the feature map serves as an information supplement to the feature map aggregated by the sub-modules of the SAS attention module, and strengthens the connection with the minor features of the feature map instead of only focusing on the salient features. This paper will combine the ResNet network with SAS attention module and ResNet network with Shuffle Attention (SA) module to experiment on the CK+ dataset, with the accuracy of 98.9836%. The results indicate that the proposed method has the potential to improve the performance of facial expression recognition.

Flower recognition has important application value in the field of flower cultivation and planting. As a kind of fine-grained image recognition, the traditional flower recognition has the problem of low recognition accuracy. In order to solve these problems, this paper proposes a neural network algorithm of ResNet structure that integrates CBAM mechanism, and adds residual blocks of attention modules to the second layer to the fifth layer of the ResNet structure. Finally, the results are output through adaptive average pooling and full connection layer. Experimental results show that the recognition accuracy of the model is close to 98% even when the recognition is difficult and there are few training data. Compared with the traditional deep learning model, the model proposed in this paper significantly improves the recognition accuracy.

Heart rate variability (HRV) is a measurement for cardiovascular health condition, and it can be obtained from photo-plethysmography signal by capturing the slight color change in human skin as a result of blood volume change. Remote photoplethysmography signals (rPPG) can be extracted from videos of human faces. Current rPPG signal postprocessing typicall y selects a certain fixed range of the frequency in the power spectrum of the signal. This paper describes a method to use signal to noise ratio to assist in the selection the adaptive range of the frequency that may contain HRV information. UBFC-rPPG dataset is used for validation. The experiment result shows a regression result of 0.88 correlation in the results from the face video compared to the ground truth data in SDRR and 0.63 in RMSSD. The proposed method of this paper has a better accuracy compared to the classical methods, to the best of our knowledge.