Search

In recent years, with the rapid development of deep learning technology, a large number of excellent convolutional neural networks (CNNs) have been proposed, many of which are based on improvements to classical methods. Based on the Inception family of methods, depthwise separable convolution was applied to Xception to achieve lightweighting, and Inception-ResNet introduces residual connections to accelerate model convergence. However, existing improvements for the Inception module often neglect further enhancement of its receptive field, while increasing the receptive field of CNNs has been widely studied and proven to be effective in improving classification performance. Motivated by this fact, three effective expansion modules are proposed in this paper. The first expansion module, Inception expand (Inception-e) module, is proposed to improve the classification accuracy by concatenating more and deeper convolutional branches. To reduce the number of parameters for Inception e, this paper proposes a second expansion module—Equivalent Inception-e (Eception) module, which is equivalent to Inception-e in terms of feature extraction capability, but which suppresses the growth of the parameter quantity brought by the expansion by effectively reducing the redundant convolutional layers; on the basis of Eception, this paper proposes a third expansion module—Lightweight Eception (Lception) module, which crosses depthwise convolution with ordinary convolution to further effectively reduce the number of parameters. The three proposed modules have been validated on the Cifar10 dataset. The experimental results show that all these extensions are effective in improving the classification accuracy of the models, and the most significant effect is the Lception module, where Lception (rank = 4) on the Cifar10 dataset improves the accuracy by 1.5% compared to the baseline model (Inception module A) by using only 0.15 M more parameters. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) need to replicate feature detectors when modeling spatial information, which reduces their efficiency. The number of replicated feature detectors or labeled training data required for such methods grows exponentially with the dimensionality of the data being used. On the other hand, space-insensitive methods are difficult to encode and express effectively due to the limitation of their rich text structures. In response to the above problems, this paper proposes a capsule network (self-attention capsule network, or SA-CapsNet) with a self-attention mechanism for text classification tasks, wherein the capsule network itself, given the feature with the symmetry hint on two ends, acts as both encoder and decoder. In order to learn long-distance dependent features in sentences and encode text information more efficiently, SA-CapsNet maps the self-attention module to the feature extraction layer of the capsule network, thereby increasing its feature extraction ability and overcoming the limitations of convolutional neural networks. In addition, in this study, in order to improve the accuracy of the model, the capsule was improved by reducing its dimension and an intermediate layer was added, enabling the model to obtain more expressive instantiation features in a given sentence. Finally, experiments were carried out on three general datasets of different sizes, namely the IMDB, MPQA, and MR datasets. The accuracy of the model on these three datasets was 84.72%, 80.31%, and 75.38%, respectively. Furthermore, compared with the benchmark algorithm, the model's performance on these datasets was promising, with an increase in accuracy of 1.08%, 0.39%, and 1.43%, respectively. This study focused on reducing the parameters of the model for various applications, such as edge and mobile applications. The experimental results show that the accuracy is still not apparently decreased by the reduced parameters. The experimental results therefore verify the effective performance of the proposed SA-CapsNet model. [ABSTRACT FROM AUTHOR]

Unmanned aerial vehicles have large prospects for organizing territory monitoring. To integrate them into this sphere, it is necessary to improve their high functionality and safety. Computer vision is one of the vital monitoring aspects. In this paper, we developed and validated a methodology for terrain classification. The overall classification procedure consists of the following steps: (1) pre-processing, (2) feature extraction, and (3) classification. For the pre-processing stage, a clustering method based on particle swarm optimization was elaborated, which helps to extract object patterns from the image. Feature extraction is conducted via Gray-Level Co-Occurrence Matrix calculation, and the output of the matrix is turned into the input for a feed-forward neural network classification stage. The developed computer vision system showed 88.7% accuracy on the selected test set. These results can provide high quality territory monitoring; prospectively, we plan to establish a self-positioning system based on computer vision. [ABSTRACT FROM AUTHOR]

Various defects are formed on the workpiece surface during the production process. Workpiece surface defects are classified according to various characteristics, which includes a bumped surface, scratched surface and pit surface. Suppliers analyze the cause of workpiece surface defects through the defect types and thus determines the subsequent processing. Therefore, the correct classification is essential regarding workpiece surface defects. In this paper, a multi-classifier decision-level fusion classification model for workpiece surface defects based on a convolutional neural network (CNN) was proposed. In the proposed model, the histogram of oriented gradient (HOG) was used to extract the features of the second fully connected layer of the CNN, and the features of the HOG were further extracted by using the local binary patterns (LBP), which was called the HOG–LBP feature extraction. Finally, this paper designed a symmetry ensemble classifier, which was used to classify the features of the last fully connected layer of the CNN and the features of the HOG–LBP. The comprehensive decision was made by fusing the classification results of the symmetry structure channels. The experiments were carried out, and the results showed that the proposed model could improve the accuracy of the workpiece surface defect classification. [ABSTRACT FROM AUTHOR]

Deep learning has been successfully applied to classification tasks in many fields due to its good performance in learning discriminative features. However, the application of deep learning to printing defect classification is very rare, and there is almost no research on the classification method for printing defects with imbalanced samples. In this paper, we present a deep convolutional neural network model to extract deep features directly from printed image defects. Furthermore, considering the asymmetry in the number of different types of defect samples—that is, the number of different kinds of defect samples is unbalanced—seven types of over-sampling methods were investigated to determine the best method. To verify the practical applications of the proposed deep model and the effectiveness of the extracted features, a large dataset of printing detect samples was built. All samples were collected from practical printing products in the factory. The dataset includes a coarse-grained dataset with four types of printing samples and a fine-grained dataset with eleven types of printing samples. The experimental results show that the proposed deep model achieves a 96.86% classification accuracy rate on the coarse-grained dataset without adopting over-sampling, which is the highest accuracy compared to the well-known deep models based on transfer learning. Moreover, by adopting the proposed deep model combined with the SVM-SMOTE over-sampling method, the accuracy rate is improved by more than 20% in the fine-grained dataset compared to the method without over-sampling. [ABSTRACT FROM AUTHOR]

Analysis of motion symmetry constitutes an important area with many applications in engineering, robotics, neurology and biomedicine. This paper presents the use of microelectromechanical sensors (MEMS), including accelerometers and gyrometers, to acquire data via mobile devices so as to monitor physical activities and their irregularities. Special attention is devoted to the analysis of the symmetry of the motion of the body when the same exercises are performed by the right and the left limb. The analyzed data include the motion of the legs on a home exercise bike under different levels of load. The method is based on signal analysis using the discrete wavelet transform and the evaluation of signal segment features such as the relative energy at selected decomposition levels. The subsequent classification of the evaluated features is performed by k-nearest neighbours, a Bayesian approach, a support vector machine, and neural networks. The highest average classification accuracy attained is 91.0% and the lowest mean cross-validation error is 0.091, resulting from the use of a neural network. This paper presents the advantages of the use of simple sensors, their combination and intelligent data processing for the numerical evaluation of motion features in the rehabilitation and monitoring of physical activities. [ABSTRACT FROM AUTHOR]

Modeling low level features to high level semantics in medical imaging is an important aspect in filtering anatomy objects. Bag of Visual Words (BOVW) representations have been proven effective to model these low level features to mid level representations. Convolutional neural nets are learning systems that can automatically extract high-quality representations from raw images. However, their deployment in the medical field is still a bit challenging due to the lack of training data. In this paper, learned features that are obtained by training convolutional neural networks are compared with our proposed hand-crafted HSIFT features. The HSIFT feature is a symmetric fusion of a Harris corner detector and the Scale Invariance Transform process (SIFT) with BOVW representation. The SIFT process is enhanced as well as the classification technique by adopting bagging with a surrogate split method. Quantitative evaluation shows that our proposed hand-crafted HSIFT feature outperforms the learned features from convolutional neural networks in discriminating anatomy image classes. [ABSTRACT FROM AUTHOR]

Classifying fine-grained categories (e.g., bird species, car, and aircraft types) is a crucial problem in image understanding and is difficult due to intra-class and inter-class variance. Most of the existing fine-grained approaches individually utilize various parts and local information of objects to improve the classification accuracy but neglect the mechanism of the feature fusion between the object (global) and object's parts (local) to reinforce fine-grained features. In this paper, we present a novel framework, namely object–part registration–fusion Net (OR-Net), which considers the mechanism of registration and fusion between an object (global) and its parts' (local) features for fine-grained classification. Our model learns the fine-grained features from the object of global and local regions and fuses these features with the registration mechanism to reinforce each region's characteristics in the feature maps. Precisely, OR-Net consists of: (1) a multi-stream feature extraction net, which generates features with global and various local regions of objects; (2) a registration–fusion feature module calculates the dimension and location relationships between global (object) regions and local (parts) regions to generate the registration information and fuses the local features into the global features with registration information to generate the fine-grained feature. Experiments execute symmetric GPU devices with symmetric mini-batch to verify that OR-Net surpasses the state-of-the-art approaches on CUB-200-2011 (Birds), Stanford-Cars, and Stanford-Aircraft datasets. [ABSTRACT FROM AUTHOR]

In the development of human–machine interfaces, facial expression analysis has attracted considerable attention, as it provides a natural and efficient way of communication. Congruence between facial and behavioral inference in face processing is considered a serious challenge that needs to be solved in the near future. Automatic facial expression is a difficult classification issue because of the high interclass variability caused by the significant interdependence of the environmental conditions on the face appearance caused by head pose, scale, and illumination occlusions from their variances. In this paper, an adaptive model for smile classification is suggested that integrates a row-transform-based feature extraction algorithm and a cascade classifier to increase the precision of facial recognition. We suggest a histogram-based cascade smile classification method utilizing different facial features. The candidate feature set was designed based on the first-order histogram probability, and a cascade classifier with a variety of parameters was used at the classification stage. Row transformation is used to exclude any unnecessary coefficients in a vector, thereby enhancing the discriminatory capacity of the extracted features and reducing the sophistication of the calculations. Cascading gives the opportunity to train an extremely precise classification by taking a weighted average of poor learners' decisions. Through accumulating positive and negative images of a single object, this algorithm can build a complete classifier capable of classifying different smiles in a limited amount of time (near real time) and with a high level of precision (92.2–98.8%) as opposed to other algorithms by large margins (5% compared with traditional neural network and 2% compared with Deep Neural Network based methods). [ABSTRACT FROM AUTHOR]

Person re-identification is the task of matching pedestrian images across a network of non-overlapping camera views. It poses aggregated challenges resulted from random human pose, clutter from the background, illumination variations, and other factors. There has been a vast number of studies in recent years with promising success. However, key challenges have not been adequately addressed and continue to result in sub-optimal performance. Attention-based person re-identification gains more popularity in identifying discriminatory features from person images. Its potential in terms of extracting features common to a pair of person images across the feature extraction pipeline has not been be fully exploited. In this paper, we propose a novel attention-based Siamese network driven by a mutual-attention module decomposed into spatial and channel components. The proposed mutual-attention module not only leads feature extraction to the discriminative part of individual images, but also fuses mutual features symmetrically across pairs of person images to get informative regions common to both input images. Our model simultaneously learns feature embedding for discriminative cues and the similarity measure. The proposed model is optimized with multi-task loss, namely classification and verification loss. It is further optimized by a learnable mutual-attention module to facilitate an efficient and adaptive learning. The proposed model is thoroughly evaluated on extensively used large-scale datasets, Market-1501 and Duke-MTMC-ReID. Our experimental results show competitive results with the state-of-the-art works and the effectiveness of the mutual-attention module. [ABSTRACT FROM AUTHOR]

To improve the recognition accuracy of coal gangue images with the back propagation (BP) neural network, a coal gangue image recognition method based on BP neural network and ASGS-CWOA (ASGS-CWOA-BP) was proposed, which makes two key contributions. Firstly, a new feature extraction method for the unique features of coal and gangue images is proposed, known as "Encircle–City Feature". Additionally, a method that applied ASGS-CWOA to optimize the parameters of the BP neural network was introduced to address to the issue of its low accuracy in coal gangue image recognition, and a BP neural network with a simple structure and reduced computational consumption was designed. The experimental results showed that the proposed method outperformed the other six comparison methods, with recognition of 95.47% and 94.37% in the training set and the test set, respectively, showing good symmetry. [ABSTRACT FROM AUTHOR]

Intrusion detection system (IDS) plays a significant role in preventing network attacks and plays a vital role in the field of national security. At present, the existing intrusion detection methods are generally based on traditional machine learning models, such as random forest and decision tree, but they rely heavily on artificial feature extraction and have relatively low accuracy. To solve the problems of feature extraction and low detection accuracy in intrusion detection, an intrusion detection model SAAE-DNN, based on stacked autoencoder (SAE), attention mechanism and deep neural network (DNN), is proposed. The SAE represents data with a latent layer, and the attention mechanism enables the network to obtain the key features of intrusion detection. The trained SAAE encoder can not only automatically extract features, but also initialize the weights of DNN potential layers to improve the detection accuracy of DNN. We evaluate the performance of SAAE-DNN in binary-classification and multi-classification on an NSL-KDD dataset. The SAAE-DNN model can detect normally and attack symmetrically, with an accuracy of 87.74% and 82.14% (binary-classification and multi-classification), which is higher than that of machine learning methods such as random forest and decision tree. The experimental results show that the model has a better performance than other comparison methods. [ABSTRACT FROM AUTHOR]

Several facial expression recognition methods cluster facial elements according to similarity and weight them considering the importance of each element in classification. However, these methods are limited by the pre-definitions of units restricting modification of the structure during optimization. This study proposes a modified support vector machine classifier called Grid Map, which is combined with reinforcement learning to improve the classification accuracy. To optimize training, the input image size is normalized according to the cascade rules of a pre-processing detector, and the regional weights are assigned by an adaptive cell size that divides each region of the image using bounding grids. Reducing the size of the bounding grid reduces the area used for feature extraction, allowing more detailed weighted features to be extracted. Error-correcting output codes with a histogram of gradient is selected as the classification method via an experiment to determine the optimal feature and classifier selection. The proposed method is formulated into a decision process and solved via Q-learning. To classify seven emotions, the proposed method exhibits accuracies of 96.36% and 98.47% for four databases and Extended Cohn-–Kanade Dataset (CK+), respectively. Compared to the basic method exhibiting a similar accuracy, the proposed method requires 68.81% fewer features and only 66.33% of the processing time. [ABSTRACT FROM AUTHOR]

Facial Expression Recognition (FER), as the primary processing method for non-verbal intentions, is an important and promising field of computer vision and artificial intelligence, and one of the subject areas of symmetry. This survey is a comprehensive and structured overview of recent advances in FER. We first categorise the existing FER methods into two main groups, i.e., conventional approaches and deep learning-based approaches. Methodologically, to highlight the differences and similarities, we propose a general framework of a conventional FER approach and review the possible technologies that can be employed in each component. As for deep learning-based methods, four kinds of neural network-based state-of-the-art FER approaches are presented and analysed. Besides, we introduce seventeen commonly used FER datasets and summarise four FER-related elements of datasets that may influence the choosing and processing of FER approaches. Evaluation methods and metrics are given in the later part to show how to assess FER algorithms, along with subsequent performance comparisons of different FER approaches on the benchmark datasets. At the end of the survey, we present some challenges and opportunities that need to be addressed in future. [ABSTRACT FROM AUTHOR]