Search

Copyright of Journal of Computer Engineering & Applications is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Abstract In the field of brain-computer interface (BCI) based on motor imagery (MI), multi-channel electroencephalography (EEG) data is commonly utilized for MI task recognition to achieve sensory compensation or precise human-computer interaction. However, individual physiological differences, environmental variations, or redundant information and noise in certain channels can pose challenges and impact the performance of BCI systems. In this study, we introduce a channel selection method utilizing Hybrid-Recursive Feature Elimination (H-RFE) combined with residual graph neural networks for MI recognition. This channel selection method employs a recursive feature elimination strategy and integrates three classification methods, namely random forest, gradient boosting, and logistic regression, as evaluators for adaptive channel selection tailored to specific subjects. To fully exploit the spatiotemporal information of multi-channel EEG, this study employed a graph neural network embedded with residual blocks to achieve precise recognition of motor imagery. We conducted algorithm testing using the SHU dataset and the PhysioNet dataset. Experimental results show that on the SHU dataset, utilizing 73.44% of the total channels, the cross-session MI recognition accuracy is 90.03%. Similarly, in the PhysioNet dataset, using 72.5% of the channel data, the classification result also reaches 93.99%. Compared to traditional strategies such as selecting three specific channels, correlation-based channel selection, mutual information-based channel selection, and adaptive channel selection based on Pearson coefficients and spatial positions, the proposed method improved classification accuracy by 34.64%, 10.8%, 3.25% and 2.88% on the SHU dataset, and by 46.96%, 5.04%, 5.81% and 2.32% on the PhysioNet dataset, respectively.

In the field of brain-computer interface (BCI) based on motor imagery (MI), multi-channel electroencephalography (EEG) data is commonly utilized for MI task recognition to achieve sensory compensation or precise human-computer interaction. However, individual physiological differences, environmental variations, or redundant information and noise in certain channels can pose challenges and impact the performance of BCI systems. In this study, we introduce a channel selection method utilizing Hybrid-Recursive Feature Elimination (H-RFE) combined with residual graph neural networks for MI recognition. This channel selection method employs a recursive feature elimination strategy and integrates three classification methods, namely random forest, gradient boosting, and logistic regression, as evaluators for adaptive channel selection tailored to specific subjects. To fully exploit the spatiotemporal information of multi-channel EEG, this study employed a graph neural network embedded with residual blocks to achieve precise recognition of motor imagery. We conducted algorithm testing using the SHU dataset and the PhysioNet dataset. Experimental results show that on the SHU dataset, utilizing 73.44% of the total channels, the cross-session MI recognition accuracy is 90.03%. Similarly, in the PhysioNet dataset, using 72.5% of the channel data, the classification result also reaches 93.99%. Compared to traditional strategies such as selecting three specific channels, correlation-based channel selection, mutual information-based channel selection, and adaptive channel selection based on Pearson coefficients and spatial positions, the proposed method improved classification accuracy by 34.64%, 10.8%, 3.25% and 2.88% on the SHU dataset, and by 46.96%, 5.04%, 5.81% and 2.32% on the PhysioNet dataset, respectively. [ABSTRACT FROM AUTHOR]

Using object-based deep learning for the urban land cover classification has become a mainstream method. This study proposed an urban land cover classification method based on segments’ object features, deep features, and spatial association features. The proposed method used the synthetic semivariance function to determine the hyperparameters of the superpixel segmentation and subsequently optimized the image superpixel segmentation result. A convolutional neural network and a graph convolutional neural network were used to obtain segments’ deep features and spatial association features, respectively. The random forest algorithm was used to classify segments based on the multidimension features. The results showed that the image superpixel segmentation results had the significant impact on the classification results. Compared with the pixel-based method, the segment-based methods generally yielded the higher classification accuracy. The strategy of multidimension feature fusion can combine the advantages of each single-dimension feature to improve the classification accuracy. The proposed method utilizing multidimension features was more efficient than traditional methods used for the urban land cover classification. The fusion of segments’ object features, deep features, and spatial association features was the best solution for achieving the urban land cover classification.

In this study, we present a novel graph-based methodology for an accurate classification of cardiac arrhythmia diseases using a single-lead electrocardiogram (ECG). The proposed approach employs the visibility graph technique to generate graphs from time signals. Subsequently, informative features are extracted from each graph and then fed into classifiers to match the input ECG signal with the appropriate target arrhythmia class. The six target classes in this study are normal (N), left bundle branch block (LBBB), right bundle branch block (RBBB), premature ventricular contraction (PVC), atrial premature contraction (A), and fusion (F) beats. Three classification models were explored, including graph convolutional neural network (GCN), multi-layer perceptron (MLP), and random forest (RF). ECG recordings from the MIT-BIH arrhythmia database were utilized to train and evaluate these classifiers. The results indicate that the multi-layer perceptron model attains the highest performance, showcasing an average accuracy of 99.02%. Following closely, the random forest achieves a strong performance as well, with an accuracy of 98.94% while providing critical intuitions.

Multi-channel electroencephalography (EEG) is used to capture features associated with motor imagery (MI) based brain-computer interface (BCI) with a wide spatial coverage across the scalp. However, redundant EEG channels are not conducive to improving BCI performance. Therefore, removing irrelevant channels can help improve the classification performance of BCI systems. We present a new method for identifying relevant EEG channels. Our method is based on the assumption that useful channels share related information and that this can be measured by inter-channel connectivity. Specifically, we treat all candidate EEG channels as a graph and define channel selection as the problem of node classification on a graph. Then we design a graph convolutional neural network (GCN) model for channels classification. Channels are selected based on the outputs of our GCN model. We evaluate our proposed GCN-based channel selection (GCN-CS) method on three MI datasets. On three datasets, GCN-CS achieves performance improvements by reducing the number of channels. Specifically, we achieve classification accuracies of 79.76% on Dataset 1, 89.14% on Dataset 2 and 87.96% on Dataset 3, which outperform competing methods significantly. [ABSTRACT FROM AUTHOR]

Recently, Graph Convolutional neural Networks (GCNs) have attracted much attention by generalizing convolutional neural networks to graph data, which includes redefining convolution and pooling operations on graphs. Due to the limitation that graph data can only focus on dyadic relations, it cannot perform well in real practice. In contrast, a hypergraph can capture high-order data interaction and is easy to deal with complex data representation using its flexible hyperedges. However, the existing methods for hypergraph convolutional networks are still not mature, and there is no effective operation for hypergraph pooling currently. Therefore, a hypergraph pooling network with a self-attention mechanism is proposed. Using a hypergraph structure for data modeling, this model can learn node hidden features with high-order data information through hypergraph convolution operation which introduces a selfattention mechanism, select important nodes both on structure and content through hypergraph pooling operation, and then obtain more accurate hypergraph representation. Experiments on text classification, dish classification, and protein classification tasks show that the proposed method outperforms recent state-of-the-art methods. [ABSTRACT FROM AUTHOR]

This study combines the present stage of the node classification problem with the fact that there is frequent noise in the graph structure of the graph convolution calculation, which can lead to the omission of some of the actual edge relations between nodes and the appearance of numerous isolated nodes. In this paper, we propose the graph neural network model ENode-GAT for improving the accuracy of small sample node classification using the method of external referencing of similar word nodes, combined with Graph Convolutional Neural Network (GCN), Graph Attention Network (GAT), and the early stop algorithm. In order to demonstrate the applicability of the model, this paper employs two distinct types of node datasets for its investigations. The first is the Cora dataset, which is widely used in node classification at this time, and the second is a small-sample Stock dataset created by Eastern Fortune's stock prospectus of the Science and Technology Board (STB). The experimental results demonstrate that the ENode-GAT model proposed in this paper obtains 85.1% classification accuracy on the Cora dataset and 85.3% classification accuracy on the Stock dataset, with certain classification advantages. It also verifies the future applicability of the model to the fields of stock classification, tender document classification, news classification, and government announcement classification, among others. [ABSTRACT FROM AUTHOR]

In recent years, overlapping entity relation extraction has received a great deal of attention and has made good progress in English. However, the research on overlapping entity relation extraction in Chinese still faces two key problems: one is the lack of datasets with overlapping entity instances, and the other is the lack of a neural network model that can effectively solve overlapping entity relation extraction. To address the above problems, this paper produces an interpersonal relationship dataset, NewsPer, for news texts and proposes a Chinese overlapping entity relation extraction model, DepCasRel. First, the model uses "Word-label" to incorporate the character features of Chinese text into the dependency analysis graph, and then uses the same binary labeling method to label the head and tail entities embedded in the text. Finally, the text's triples are extracted. DepCasRel solves the problem that traditional methods make it difficult to extract triples with overlapping entities. Experiments on our manually annotated dataset NewsPer show that DepCasRel can effectively encode the semantic and structural information of text and improve the performance of an overlapping entity relation extraction model. [ABSTRACT FROM AUTHOR]

Deep learning technology, such as fully convolutional networks (FCNs), have shown competitive performance in the automatic extraction of buildings from high-resolution aerial images (HRAIs). However, there are problems of over-segmentation and internal cavity in traditional FCNs used for building extraction. To address these issues, this paper proposes a new building graph convolutional network (BGC-Net), which optimizes the segmentation results by introducing the graph convolutional network (GCN). The core of BGC-Net includes two major modules. One is an atrous attention pyramid (AAP) module, obtained by fusing the attention mechanism and atrous convolution, which improves the performance of the model in extracting multi-scale buildings through multi-scale feature fusion; the other is a dual graph convolutional (DGN) module, the build of which is based on GCN, which improves the segmentation accuracy of object edges by adding long-range contextual information. The performance of BGC-Net is tested on two high spatial resolution datasets (Wuhan University building dataset and a Chinese typical city building dataset) and compared with several state-of-the-art networks. Experimental results demonstrate that the proposed method outperforms several state-of-the-art approaches (FCN8s, DANet, SegNet, U-Net, ARC-Net, BAR-Net) in both visual interpretation and quantitative evaluations. The BGC-Net proposed in this paper has better results when extracting the completeness of buildings, including boundary segmentation accuracy, and shows great potential in high-precision remote sensing mapping applications. [ABSTRACT FROM AUTHOR]

Mobile-Edge Computing (MEC) is an emerging paradigm in the Internet of Vehicles (IoV) to meet the ever-increasing computation demands of smart applications. To provide satisfactory computation performance, it is of significant importance to conduct computation offloading in IoV. In this paper, we investigate a multi-vehicle IoV system assisted by MECs with limited computation resources, where vehicles with complex applications can offload their subtasks to MEC servers. Applications are modeled as interdependent subtasks with general random task graphs, different from existing works with independent ones. To maximize the average logarithmic data processing rate (LDPR), the computation offloading problem is formulated as a time-average optimization with long-term constraints, which results from variable vehicle number, various applications and time-varying communication channels. To reduce the cooperation overhead, we propose a multi-agent Proximal Policy Optimization algorithm (Ly-MAPPO) which requires local observation only to solve the subproblems achieved by Lyapunov optimization technique in real time. In addition, to improve the performance of the Ly-MAPPO algorithm, Graph Convolutional Neural Network (GCN) is introduced to extract inter-dependencies between subtasks. Extensive simulations show that the GCN embedded Ly-MAPPO outperforms other baseline algorithms, e.g., greedy algorithm and gene algorithm, etc., for different traffic loads and computation resources in MEC servers. [ABSTRACT FROM AUTHOR]

This brief presents the first IC implementation of graph convolutional neural network (GCN) accelerator chip. A sparsity aware dataflow optimized for sub-block-wise processing of three different matrices in GCN is proposed to improve the utilization ratio of computing resources while reducing the amount of redundant access of off-chip memory. The implemented accelerator in 28-nm CMOS produces 384G NZ outputs/J for the extremely sparse matrix multiplications of the GCN. It shows 58k-to-143k, 38k-to-92k and 5k-to-13k Graph/J for the benchmark graph datasets of Cora, Citeseer and Pubmed, respectively. The energy efficiency in Graph/J of the proposed 16b ASIC implementation shows about 4-to- $11\mathbf {\times }$ and 8-to- $25\mathbf {\times }$ improvements compared to the previously reported 8b FPGA and 32b FPGA implementations, respectively. [ABSTRACT FROM AUTHOR]

Sign language is the main channel for hearing-impaired people to communicate with others. It is a visual language that conveys highly structured components of manual and non-manual parameters such that it needs a lot of effort to master by hearing people. Sign language recognition aims to facilitate this mastering difficulty and bridge the communication gap between hearing-impaired people and others. This study presents an efficient architecture for sign language recognition based on a convolutional graph neural network (GCN). The presented architecture consists of a few separable 3DGCN layers, which are enhanced by a spatial attention mechanism. The limited number of layers in the proposed architecture enables it to avoid the common over-smoothing problem in deep graph neural networks. Furthermore, the attention mechanism enhances the spatial context representation of the gestures. The proposed architecture is evaluated on different datasets and shows outstanding results. [ABSTRACT FROM AUTHOR]

This study combines the present stage of the node classification problem with the fact that there is frequent noise in the graph structure of the graph convolution calculation, which can lead to the omission of some of the actual edge relations between nodes and the appearance of numerous isolated nodes. In this paper, we propose the graph neural network model ENode-GAT for improving the accuracy of small sample node classification using the method of external referencing of similar word nodes, combined with Graph Convolutional Neural Network (GCN), Graph Attention Network (GAT), and the early stop algorithm. In order to demonstrate the applicability of the model, this paper employs two distinct types of node datasets for its investigations. The first is the Cora dataset, which is widely used in node classification at this time, and the second is a small-sample Stock dataset created by Eastern Fortune’s stock prospectus of the Science and Technology Board (STB). The experimental results demonstrate that the ENode-GAT model proposed in this paper obtains 85.1% classification accuracy on the Cora dataset and 85.3% classification accuracy on the Stock dataset, with certain classification advantages. It also verifies the future applicability of the model to the fields of stock classification, tender document classification, news classification, and government announcement classification, among others.

Recognition accuracy and response time are both critically essential ahead of building the practical electroencephalography (EEG)-based brain–computer interface (BCI). However, recent approaches have compromised either the classification accuracy or the responding time. This paper presents a novel deep learning approach designed toward both remarkably accurate and responsive motor imagery (MI) recognition based on scalp EEG. Bidirectional long short-term memory (BiLSTM) with the attention mechanism is employed, and the graph convolutional neural network (GCN) promotes the decoding performance by cooperating with the topological structure of features, which are estimated from the overall data. Particularly, this method is trained and tested on the short EEG recording with only 0.4 s in length, and the result has shown effective and efficient prediction based on individual and groupwise training, with 98.81% and 94.64% accuracy, respectively, which outperformed all the state-of-the-art studies. The introduced deep feature mining approach can precisely recognize human motion intents from raw and almost-instant EEG signals, which paves the road to translate the EEG-based MI recognition to practical BCI systems.

Deep learning technology, such as fully convolutional networks (FCNs), have shown competitive performance in the automatic extraction of buildings from high-resolution aerial images (HRAIs). However, there are problems of over-segmentation and internal cavity in traditional FCNs used for building extraction. To address these issues, this paper proposes a new building graph convolutional network (BGC-Net), which optimizes the segmentation results by introducing the graph convolutional network (GCN). The core of BGC-Net includes two major modules. One is an atrous attention pyramid (AAP) module, obtained by fusing the attention mechanism and atrous convolution, which improves the performance of the model in extracting multi-scale buildings through multi-scale feature fusion; the other is a dual graph convolutional (DGN) module, the build of which is based on GCN, which improves the segmentation accuracy of object edges by adding long-range contextual information. The performance of BGC-Net is tested on two high spatial resolution datasets (Wuhan University building dataset and a Chinese typical city building dataset) and compared with several state-of-the-art networks. Experimental results demonstrate that the proposed method outperforms several state-of-the-art approaches (FCN8s, DANet, SegNet, U-Net, ARC-Net, BAR-Net) in both visual interpretation and quantitative evaluations. The BGC-Net proposed in this paper has better results when extracting the completeness of buildings, including boundary segmentation accuracy, and shows great potential in high-precision remote sensing mapping applications.

• The present study proposed a novel graph convolution model to forecast future traffic speeds. • The proposed model differentiated the intensity of connecting to neighbor roads unlike existing GCNs. • The present study was focused on devising a GCN model that mimic true propagation patterns of traffic. • The proposed model shows promise for application to a real-time traffic forecasting system. The traffic state in an urban transportation network is determined via spatio-temporal traffic propagation. In early traffic forecasting studies, time-series models were adopted to accommodate autocorrelations between traffic states. The incorporation of spatial correlations into the forecasting of traffic states, however, involved a computational burden. Deep learning technologies were recently introduced to traffic forecasting in order to accommodate the spatio-temporal dependencies among traffic states. In the present study, we devised a novel graph-based neural network that expanded the existing graph convolutional neural network (GCN). The proposed model allowed us to differentiate the intensity of connecting to neighbor roads, unlike existing GCNs that give equal weight to each neighbor road. A plausible model architecture that mimicked real traffic propagation was established based on the graph convolution. The domain knowledge was efficiently incorporated into a neural network architecture. The present study also employed a generative adversarial framework to ensure that a forecasted traffic state could be as realistic as possible considering the joint probabilistic density of real traffic states. The forecasting performance of the proposed model surpassed that of the original GCN model, and the estimated adjacency matrices revealed the hidden nature of real traffic propagation. [ABSTRACT FROM AUTHOR]

In recent years, object detection has shown excellent results on a large number of annotated data, but when there is a discrepancy between the annotated data and the real test data, the performance of the trained object detection model is often degraded when it is directly transferred to the real test dataset. Compared with natural images, remote sensing images have great differences in appearance and quality. Traditional methods need to re-label all image data before interpretation, which will consume a lot of manpower and time. Therefore, it is of practical significance to study the Cross-Domain Adaptation Object Detection (CDAOD) of remote sensing images. To solve the above problems, our paper proposes a Rotation-Invariant and Relation-Aware (RIRA) CDAOD network. We trained the network at the image-level and the prototype-level based on a relation aware graph to align the feature distribution and added the rotation-invariant regularizer to deal with the rotation diversity. The Faster R-CNN network was adopted as the backbone framework of the network. We conducted experiments on two typical remote sensing building detection datasets, and set three domain adaptation scenarios: WHU 2012 → WHU 2016, Inria (Chicago) → Inria (Austin), and WHU 2012 → Inria (Austin). The results show that our method can effectively improve the detection effect in the target domain, and outperform competing methods by obtaining optimal results in all three scenarios.

Although the deep semantic segmentation network (DSSN) has been widely used in remote sensing (RS) image semantic segmentation, it still does not fully mind the spatial relationship cues between objects when extracting deep visual features through convolutional filters and pooling layers. In fact, the spatial distribution between objects from different classes has a strong correlation characteristic. For example, buildings tend to be close to roads. In view of the strong appearance extraction ability of DSSN and the powerful topological relationship modeling capability of the graph convolutional neural network (GCN), a DSSN-GCN framework, which combines the advantages of DSSN and GCN, is proposed in this paper for RS image semantic segmentation. To lift the appearance extraction ability, this paper proposes a new DSSN called the attention residual U-shaped network (AttResUNet), which leverages residual blocks to encode feature maps and the attention module to refine the features. As far as GCN, the graph is built, where graph nodes are denoted by the superpixels and the graph weight is calculated by considering the spectral information and spatial information of the nodes. The AttResUNet is trained to extract the high-level features to initialize the graph nodes. Then the GCN combines features and spatial relationships between nodes to conduct classification. It is worth noting that the usage of spatial relationship knowledge boosts the performance and robustness of the classification module. In addition, benefiting from modeling GCN on the superpixel level, the boundaries of objects are restored to a certain extent and there are less pixel-level noises in the final classification result. Extensive experiments on two publicly open datasets show that DSSN-GCN model outperforms the competitive baseline (i.e., the DSSN model) and the DSSN-GCN when adopting AttResUNet achieves the best performance, which demonstrates the advance of our method.

In recent years, object detection has shown excellent results on a large number of annotated data, but when there is a discrepancy between the annotated data and the real test data, the performance of the trained object detection model is often degraded when it is directly transferred to the real test dataset. Compared with natural images, remote sensing images have great differences in appearance and quality. Traditional methods need to re-label all image data before interpretation, which will consume a lot of manpower and time. Therefore, it is of practical significance to study the Cross-Domain Adaptation Object Detection (CDAOD) of remote sensing images. To solve the above problems, our paper proposes a Rotation-Invariant and Relation-Aware (RIRA) CDAOD network. We trained the network at the image-level and the prototype-level based on a relation aware graph to align the feature distribution and added the rotation-invariant regularizer to deal with the rotation diversity. The Faster R-CNN network was adopted as the backbone framework of the network. We conducted experiments on two typical remote sensing building detection datasets, and set three domain adaptation scenarios: WHU 2012 → WHU 2016, Inria (Chicago) → Inria (Austin), and WHU 2012 → Inria (Austin). The results show that our method can effectively improve the detection effect in the target domain, and outperform competing methods by obtaining optimal results in all three scenarios. [ABSTRACT FROM AUTHOR]

Although the deep semantic segmentation network (DSSN) has been widely used in remote sensing (RS) image semantic segmentation, it still does not fully mind the spatial relationship cues between objects when extracting deep visual features through convolutional filters and pooling layers. In fact, the spatial distribution between objects from different classes has a strong correlation characteristic. For example, buildings tend to be close to roads. In view of the strong appearance extraction ability of DSSN and the powerful topological relationship modeling capability of the graph convolutional neural network (GCN), a DSSN-GCN framework, which combines the advantages of DSSN and GCN, is proposed in this paper for RS image semantic segmentation. To lift the appearance extraction ability, this paper proposes a new DSSN called the attention residual U-shaped network (AttResUNet), which leverages residual blocks to encode feature maps and the attention module to refine the features. As far as GCN, the graph is built, where graph nodes are denoted by the superpixels and the graph weight is calculated by considering the spectral information and spatial information of the nodes. The AttResUNet is trained to extract the high-level features to initialize the graph nodes. Then the GCN combines features and spatial relationships between nodes to conduct classification. It is worth noting that the usage of spatial relationship knowledge boosts the performance and robustness of the classification module. In addition, benefiting from modeling GCN on the superpixel level, the boundaries of objects are restored to a certain extent and there are less pixel-level noises in the final classification result. Extensive experiments on two publicly open datasets show that DSSN-GCN model outperforms the competitive baseline (i.e., the DSSN model) and the DSSN-GCN when adopting AttResUNet achieves the best performance, which demonstrates the advance of our method. [ABSTRACT FROM AUTHOR]