Search

In order to solve the problem that it is difficult for the model to perceive implicit information due to the lack of external knowledge in rumor detection, which limits the ability of the model to mine deep information, this paper proposed knowledge graph based multi-feature fusion rumor detection (KGMRD) method. Firstly, for each event, it constructed posts and comments together into a text sequence and used a classifier to extract the emotional features. This paper constructed a knowledge graph based on text using ConceptNet and aggregated the entity representation in the knowledge graph with the semantic features of text using the attention mechanism, so as to obtain the enhanced semantic feature representation. Secondly, in terms of communication structure, for each event, this paper built its communication structure diagram based on the propagation and forwarding relationship of the post, and used DropEdge to prune the communication structure diagram, so as to obtain more effective communication structure characteristics. Finally, it fused the obtained features to get a new representation and compared seven models including SVM-RBF on three real datasets of Weibo, Twitter15 and Twitter16. The experimental results show that compared with the current baseline with the best effect, the KGMRD method has the best ACC on the Weibo dataset and improves the ACC by 1.1%, and there is a 2.2% improvement on Twitter15 and Twitter16 dataset in ACC. The experiment proves that the KGMRD method is reasonable and effective. [ABSTRACT FROM AUTHOR]

U-Net has achieved great success in the field of image segmentation. However, some of the position information is lost in the process of convolution and downsampling, model is difficult to learn global and long-range semantic interaction information and lacks the ability to integrate global and local information. To extract rich local detail and contextual information, this paper proposed an image segmentation network called MLFCNet, combining a convolutional module and a capsule encoder. Based on the U-Net, this paper introduced a capsule network module to learn target positional information and the relationships between local and global information. At the same time, the proposed attention mechanism could retain the information discarded by the network pooling layer. This paper designed a new attention mechanism so that multi-scale features could be fused, where global information was captured and background noise was suppressed. In addition, it proposed a new local feature co-occurrence algorithm to better learn the relationship between local features. The proposed method was compared with nine methods on two public datasets, mIoU improves 4.7% and Dice coefficient improves 1.7% in liver medical images compared to the second highest performing model. Experimental results on the dataset of liver and dataset of human show that under the same experimental conditions, the proposed network is superior to U-Net and other mainstream image segmentation networks. [ABSTRACT FROM AUTHOR]

Addressing the shortcomings of existing knowledge completion models based on graph neural networks in handling the heterogeneity of knowledge graphs and the limitations posed by most models’ adoption of a single static entity representation, this paper introduces a model based on the graph attention mechanism for disentangled neighborhood information aggregation. Initially, this paper learns the disentangled representation of each entity,providing a multi-component representation for the latent influential factors of entities. Using the attention mechanism,the model selects the most influential latent factors as connection elements for two connected entities. Subsequently,by leveraging a relation-aware attention mechanism, the model adaptively aggregates neighborhood messages at the entity factor level, effectively reducing interference from irrelevant information during aggregation and significantly enhancing the model’s semantic representation capability. Moreover, to focus on the most relevant entity component in the scoring process with a given relation, this paper introduces an adaptive scoring coefficient, enabling the model to perceive the relevance between the given relationship and various entity components adaptively. Experimental results on the WN18RR and FB15K-237 datasets indicate that the proposed model outperforms other advanced baseline models in knowledge graph completion tasks, substantially enhancing the model’s expressive power. [ABSTRACT FROM AUTHOR]

With the development of deep learning, object detection has gradually replaced manual quality inspection as the primary direction for industrial quality control. However, in the complex texture of the tile surface, there are more lowvisibility small defects, and the interference from the complex textured background is serious. This results in a high false detection and false alarm rate using traditional object detection methods. To enhance the efficiency of defect detection, this paper proposed a defect detection approach on complex textured tile surfaces based on the joint attention mechanisms of channels and spatial. Firstly, to enhance the feature expression of small defects, it proposed a selective feature fusion method by modeling the relationship between deep and shallow feature channels. Secondly, it designed a joint channel and spatial attention module that selects key feature channels and suppresses texture regions through channel and spatial attention,enabling the model to focus on learning defect features and enhancing its ability to discriminate between defects and texture.Finally, it validated the approach on a dataset of complexly textured ceramic tile surface defects. The experimental results demonstrate that compared to the AFF and CBAM methods, the selective feature fusion method and channel ＆ spatial joint attention achieved improvements of 5.3AP and 6.32AP, respectively. In addition, this paper compared the overall approach with the existing tile detection method YOLOv5 and texture fabric defect detection method AFAM, showing that it outperformed these methods, with respective improvements of 1.32AP and 2.12AP. [ABSTRACT FROM AUTHOR]

Mainstream fully supervised deep learning segmentation models can achieve good results when trained on abundant labeled data, but the image segmentation in the medical field faces the challenges of high annotation cost and diverse segmentation targets, often lacking sufficient labeled data. The model proposed in this paper incorporates the idea of extracting labels from data through self-supervision, utilizing superpixels to represent image characteristics for image segmentation under conditions of small sample annotation. The introduction of multiple attention mechanisms allows the model to focus more on spatial features of the image. The position attention module and channel attention module aim to fuse multi-scale features within a single image, while the external attention module highlights the connections between different samples. Experiments were conducted on the CHAOS healthy abdominal organ dataset. In the extreme case of the 1-shot, DSC reached 0.76, which is about 3%higher than the baseline result. In addition, this paper explores the significance of few-shot learning by adjusting the number of N-way-K-shot tasks. Under the 7-shot setting, DSC achieves significant improvement, which is within an acceptable range of the segmentation effect based on full supervision based on deep learning. [ABSTRACT FROM AUTHOR]

In order to alleviate the inherent information differences between different user social spaces and interest spaces in recommendation systems and the problem of ignoring high-order neighbors, this paper proposed AGCRSR. Firstly, the algorithm used a mapping matrix in the embedding layer to convert the initial feature vectors into adaptive embedding. Secondly, it intro- duced an attention mechanism to aggregate different aspects of user embeddings, and used a graph convolutional network to li- nearly learn the potential representations of users and items. Finally, it used the adaptive module to aggregate user representations and predict the final recommendation results of users for the project using the inner product function. This paper conducted comparative experiments on the LastFM and Ciao datasets and compared with other baseline algorithms. The experimental results show that the recommendation performance of the AGCRSR is significantly improved compared to other algorithms. [ABSTRACT FROM AUTHOR]

Low-light image segmentation is always the difficulty of image segmentation. The low contrast and high fuzziness caused by low light make this kind of image segmentation much more difficult than general image segmentation. In order to improve the accuracy of semantic segmentation in low light environment, this paper proposed a semantic segmentation model of low light scene with filtering light style under noise guidance(SFIS) according to the characteristics of low-light image. The model comprehensively used signal-to-noise ratio as prior knowledge, and adopted different distance interaction for different noise regions in the image by guiding the self-attention operation in the long distance branch and the feature fusion of long/short distance branches. This paper also further designed an illumination filter, which was a module that further extracted the illumination style information from the overall style of the image. By alternately training the illumination filter and the semantic segmentation model, the lighting style gap between different lighting conditions was gradually reduced, so that the segmentation network could learn illumination invariant features. The proposed model outperforms the previous work on the dataset LLRGBD and achieves the best results. The mIoU on the real dataset LLRGBD-real reaches 66.8%, it shows that the proposed long and short distance branch module and the illumination filter module can effectively improve the semantic segmentation ability of the model in low light environment. [ABSTRACT FROM AUTHOR]

The current advanced session recommendation algorithms mainly use graph neural network to mine the pairwise transformation relationships of items from the global and target sessions, and compress the target session into a fixed vector representation, ignoring the complex high-order information between items and the impact of target items on user preference diversity. To this end, this paper proposed a hypergraph convolution network and target multi-intention perception for session-based recommendation algorithm HCN-TMP. This algorithm expressed user preference by learning session representation. Firstly, it constructed a session graph based on the target session, and constructed a hypergraph based on the global session. It transformed the original item embedding representation that reflected the user's coupling intention into a multi factor embedding representation of the item through intention disentanglement technology. Then, it learned the item representations of the session level and global level of the target session node through graph attention network and hypergraph convolutional network respectively, and used the distance correlation loss function to enhance the independence between the multi-factor embedded blocks. Next, it embedded the node location information in the target session, weighted the attention weight of each node, and got the session representation of the global level and session level. It used comparative learning to maximize the mutual information of the two. Through the target multi-intention perception, it adaptively learned the multi-intention user preferences in the target session for different target items, obtained the session representation of the target perception level. Finally, it linearly fused the three level session representations to obtain the final session representation. This paper carried out the experiments on two public data sets, Tmall and Nowplaying. The experimental results verify the effectiveness of the HCN-TMP algorithm. [ABSTRACT FROM AUTHOR]

Predictive business process monitoring(PBPM) represents a vital research field within BPM that aims to accurately predict future behavioral events. At present, deep learning methods are widely used in PBPM research. However, most of these methods consider only a single event-control flow perspective and do not fuse the attribute-data flow perspective for process prediction. To address this issue, this paper proposed a method called the fusion multi-perspective(FMP) framework based on a two-layer BERT neural network. Firstly, the first layer of BERT was used to learn attribute-data flow information. Subsequently, the second layer of BERT learnt event-behavior control flow information. Finally, the FMP framework combined data flow and control flow to achieve multi-perspective process prediction. Experimental results on real event logs demonstrate that, compared to other research methods, the FPM framework yields higher accuracy in predicting the next event activity. This validates that the FPM framework, which merges multi-perspective views of processes, enables a more comprehensive and in-depth analysis of complex process behaviors while enhancing predictive performance. [ABSTRACT FROM AUTHOR]

The success of emotional dialogue systems relies on the ability to comprehend, perceive, and express emotions, while facial expressions and personality can also help. However, despite the crucial importance of this multi-modal information in multi-turn emotional dialogues, existing systems still need to be improved to leverage multi-modal information's advantages and overlook the significance of contextual relevance. To address this issue, this paper proposed a multi-turn emotional dialogue generation model based on a dua-decoding method(MEDG-DD). The model utilized a heterogeneous graph neural network encoder to integrate historical dialogue, facial expressions, emotion flow, and speaker information, obtaining a more comprehensive dialogue context. Subsequently, it employed a dual-decoding mechanism based on attention to generate emotionally rich expressions relevant to the dialogue context. Experimental results demonstrate that the proposed model effectively integrates multi-modal information, achieving more accurate, natural, and coherent emotional expressions. Compared to the traditional ReCoSa model, this model exhibits significant improvements across various evaluation metrics. [ABSTRACT FROM AUTHOR]

To address the low segmentation accuracy of multifidus muscle lesion sites in MRI images of patients with lumbar disc herniation, this paper proposes a new model to improve the U2-Net network with the goal that the encoding and decoding subnetworks are interconnected by a series of nested jump paths. To reduce the semantic missing of feature maps in the encoding and decoding subnetworks, the jump connections in the middle of RSU-7,RSU-6,RSU-5,and RSU-4 in the U2-Net model are redesigned, while the RSU-4F part remains unchanged. In addition, the channel attention module is added to enable the net to focus on channels of higher contribution to task, thus extract high quality multifractal muscle features. The experiments on the multifidus muscle MRI image dataset show that the improved U2-Net outperforms U-Net,U2-Net and U-Net++ network in indicators of Dice, HD and MIoU. It can be concluded that the proposed approach has good performance on MRI image segmentation of multifidus muscle, which can assist doctors to make diagnosis. [ABSTRACT FROM AUTHOR]

To address the issue that existing video colorization methods are complicated to guarantee both coloring quality and temporal consistency, this paper proposed a video colorization method AMVC-GAN combining attention mechanism and multi-scale feature fusion. Firstly, it proposed a GAN-based video colorization network model. It designed a multi-scale feature fusion module in the generator of GAN with a cyclic time network as the main body to obtain information of different time frequencies. Secondly, to effectively consider the relationship between adjacent frames, it used the features extracted from different time frequencies to strengthen the connection between frames as a way to enhance the temporal consistency of colorization. Finally, to obtain more helpful information, it introduced an attention module in the upsampling part, and optimally trained the results by utilizing PatchGAN to enhance the final colorization effect. Comparing with the state-of-the-art automatic video colorization methods on DAVIS and VIDEVO datasets, the results show that AMVC-GAN ranks first in multiple indicators, with better time consistency and colorization effect. Compared with other methods, AMVC-GAN can effectively reduce time flicker, while ensuring more real and natural colorization effect. [ABSTRACT FROM AUTHOR]

To address the problems of large amount of noise and lack of contextual information in short texts on the Web, this paper proposed a text sentiment analysis model based on BERT and hypergraph with dual attention mechanism. This method first utilized BERT for dynamic feature extraction of sentiment texts. Meanwhile it mined the contextual, topic and semantic dependency information of the text to model it into a hypergraph, and then aggregated the above information through the dual graph attention mechanism. Finally, it spliced the features extracted by BERT and Hypergraph with dual attention network, and obtained the prediction result after softmax layer. The accuracy of this model on the e-commerce review dataset and the microblog text dataset reaches 95.49% and 79.83% respectively,which is 2.27%-3.45% and 6.97%-11.69% higher than the baselines, respectively. The experimental results show that the model can significantly improve the accuracy of sentiment analysis for Chinese web short texts. [ABSTRACT FROM AUTHOR]

At present, most of the aspect-based sentiment analysis methods based on graph convolutional networks use the syntactic knowledge, semantic knowledge and sentiment knowledge of text to construct text dependency. However,there are few studies on the use of text word order knowledge to build text dependency, which leads to the inability of graph convolutional networks to effectively use text word order knowledge to guide aspects to learn contextual sentiment information, thus limiting its performance. Aiming at the above problems, this paper proposes an aspect-based sentiment analysis model of Dual-channel Graph Convolutional Network with Word-order Knowledge (WKDGCN), which is composed of Word-order Graph Convolutional Network (WoGCN) and the Sentiment and Attention enhanced syntactic Graph Convolutional Network (SAGCN). Specifically, WoGCN constructs a graph convolutional network based on the word-order knowledge of text, and learns contextual sentiment information by relying on the characteristics of the guiding aspect of the text word order. SAGCN uses the sentiment knowledge in SenticNet combined with the attention mechanism to enhance syntactic dependence, and uses the enhanced syntactic dependence to construct a graph convolutional network to guide the aspect features to learn contextual emotional information. Finally, the features of two graph convolutional networks are fused for sentiment classification. In addition,a weight allocation strategy is designed to keep the context weights consistent while amplifying the weights of the terms to avoid erroneous calculation of the semantic correlation between the features of the terms and the important features.Finally, experimental results on several public data sets show that the proposed method is superior to the comparison model. [ABSTRACT FROM AUTHOR]

Distant supervision relation extraction uses the automatic alignment of natural language texts and knowledge bases to generate labeled training datasets, solving the problem of manual sample labeling. In the current research, most distant supervision does not pay attention to the long-tail data, so most of the sentence bags obtained by distant supervision contain too few sentences. These sentence bags cannot truly and comprehensively express the data itself. This paper proposes a distant supervised relation extraction model (PG+PTATT) based on position-type attention mechanism and graph convolutional network. According to the similarity between sentence bags, Graph Convolutional Networks (GCN) aggregate the implicit high-level features of similar sentence bags to optimize the sentence bags and obtain more prosperous and more comprehensive feature information of the sentence bags. At the same time, an attention mechanism, Position-Type Attention (PTATT) is constructed, which can solve the problem of wrong labels in distant supervision relation extraction: using the position relationships between entity words and non-entity words and type relationships are modeled to reduce the impact of noisy words. The proposed model is experimentally verified on the dataset New York Times (NYT), and the experimental results show that the proposed model can effectively solve the problems existing in distant supervision relation extraction; and it can effectively improve the accuracy of relation extraction.. [ABSTRACT FROM AUTHOR]

To address the problem of numerous hyperparameters, loss of long time series information and difficulty in distinguishing primary and secondary features in Long Short-Term Memory network (LSTM) for grain yield ca- pacity prediction, this paper proposes a combined data-driven grain capacity forecasting model. In the hyperparameter part, the proposed model performs hyperparameter search optimization for LSTM by introducing Dy- namic Weights and Laplacian variation of Bald Eagle Search Optimization Algorithm (WLBES), to avoid the process of manual parameter adjustment. In the prediction part, the proposed model uses Ridge Regression (RR) to correct the residuals of the prediction results to make up for the deficiency of LSTM data loss, and adds an attention mechanism to distinguish primary and secondary features by weight size to enhance the importance of features with greater relevance to grain production. The results show that the combined WLBES-LSTM-RR model decreases the root mean square error (RMSE) by 75% and 19% compared with the LSTM and WLBES-LSTM models, respective- ly, and substantially decreases the RMSE compared with other combined models of optimized LSTM. This combined model has higher prediction accuracy in grain yield capacity prediction. [ABSTRACT FROM AUTHOR]

Traditional traffic flow prediction models model historical data in terms of time and space, ignoring the internal potential temporal periodicity of traffic data and the distance characteristics and similarity spatial characteristics of nodes between traffic networks. Based on this, this paper proposed a multi-channel spatio-temporal encoder model MC-STGNN for traffic flow prediction to improve the accuracy of traffic flow prediction. Firstly, it processed the traffic data into a three channel periodic time series, and encoded the overall sequence data with temporal and adaptive spatial positions to extract dynamic correlations between road network nodes. Secondly, it introduced a multi-heads self-attention mechanism with convolutional structure to capture varying degrees of temporal correlation of periodic data to a greater extent. Finally, it proposed a graph generator to generate a new spatiotemporal map, extracting similarity and distance features between road network nodes, and integrating the spatial information of the original map and the new spatiotemporal map using a gated graph convolutional network. It conducted comprehensive traffic flow prediction experiments for an hour on the highway datasets PEMS03 and PEMS08. The experimental results show that the MC-STGNN model has better performance indicators compared to other baseline models, indicating that the MC-STGNN model has better modeling ability. [ABSTRACT FROM AUTHOR]

The maximum mean discrepancy (MMD), as a representative distribution metric between source domain and target domain, has been widely applied in unsupervised domain adaptation (UDA), where both domains follow different distributions, and the labels from source domain are merely available. However, MMD and its class-wise variants possibly ignore the intra-class compactness and interclass separability, thus reducing discriminability of feature representation. This paper proposes a focally discriminative loss for unsupervised domain adaptation. This method endeavors to improve the discriminative ability of MMD from two aspects: (1) the weights are re-designed for MMD in order to align the distribution of relatively hard classes across domains; (2) a focally contrastive loss is explored to tradeoff the positive sample pairs and negative ones for better discrimination. The integration of both losses can not only make the intra-class features close, but also push away the inter-class features far from each other. Moreover, the improved loss is simple yet effective, and it can be extended to the network structure of the attention mechanism. Experiments on several domain adaptation datasets verify the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

There are quite a few comment sentences without emotional words in aspect-level emotional texts, and the study of their emotions is called aspect-level implicit sentiment analysis. The existing models have the problems that the context information related to aspect words may be lost in the pre-training process, and the deep features in the context cannot be accurately extracted. Aiming at the first problem, this paper constructs an aspect-aware BERT pre-training model, and introduces aspect words into the input embedding structure of basic BERT to generate word vectors related to aspect words. Aiming at the second problem, this paper constructs a context-aware attention mechanism. For the deep hidden vectors obtained from the coding layer, the semantic and syntactic information is introduced into the attention weight calculation process, so that the attention mechanism can more accurately assign weights to the context related to aspect words. The results of comparative experiments show that the proposed model outperforms the baseline model. BERT [ABSTRACT FROM AUTHOR]

In order to solve the problems of poor anti-interference and invasion of user privacy in traditional methods of using cameras for human activity recognition, this paper proposed a human activity recognition network based on spatiotemporal attention of millimeter wave radar 3D point cloud data to achieve accurate perception of intelligent application context. The network firstly used a secondary sliding time window to accumulate and separate the point cloud data generated by human activities as the input of the classifier, then used the PointLSTM unit to aggregate point features and states according to the point cloud coordinate relationship to extract the time sequence features of human activities, and then spliced temporal-spatial features, reduced the overall network computation and enhanced the network’s aggregation ability for local featured through sampling grouping modules, and finally used a stacked attention module to deeply fuse global and local features in temporal-spatial point cloud data to complete the accurate classification of human activities. This paper used millimeter wave radar to collect point cloud datasets of various human activities, the experimental results show that the average accuracy of the proposed spatiotemporal attention network can reach 98.64%,which can effectively identify complex and small-difference human activities, and meet the requirements of the human activity recognition system. [ABSTRACT FROM AUTHOR]

Aiming at the problem of insufficient feature extraction ability of deep clustering algorithm on multivariate time series data( MTS), This paper proposes a new deep clustering structure model (MDTC. In order to extract the key features of MTS and realize dimensionality reduction, one-dimensional convolution learns the attribute and temporal dimension feature representation of MTS and the autoencoder structure composed of network layers such as recurrent neural network; To improve the model’s ability to represent temporal features, this paper proposes MCBAM temporal attention module, which is used to enhance the representation features of different time periods in the MTS sequence. This paper conducts experiments on nine publicly available UEA multivariate time series datasets, compared with other autoencoders, the autoencoder structure of the model improves by 2%-9% on 7 data sets; Compared with other attention modules, the MCBAM module of the model improves by 0.3%-2% on the six datasets. Experiments show the effectiveness of the MDTC model structure and MCBAM module, and the model has excellent performance compared with other clustering algorithms. [ABSTRACT FROM AUTHOR]

COVID-19 is spreading rapidly around the world. In order to diagnose it quickly and accurately and thus block the chain of epidemic transmission, the study proposed a deep learning-based classification network DLDA-A-DenseNet. Firstly, DenseNet-201 combined deep layer dense aggregation to aggregate feature information at different stages to enhance its ability to identify and localize lesions. Secondly, this paper proposed efficient multi-scale long-range attention to refine the aggregated features. Moreover, this paper used a balanced sampling training strategy to eliminate the bias for the class imbalance problem of CT image dataset. Testing on the China consortium of chest CT image investigation dataset, the method improved 2.24%,3.09%,2.09%,2.60% and 3.48% in accuracy, recall, precision, F1 score and Kappa coefficient compared with DenseNet-201,and achieved an optimal accuracy of 99.50% on COVID-CISet image dataset. The results show that the proposed COVID-19 CT image classification method can fully extract the lesion features of CT slices compared with other methods, and has higher classification accuracy and good generalizability. [ABSTRACT FROM AUTHOR]

With the development of artificial intelligence, digital images have been widely used in various fields. However, due to the appearance of image editing software, a large number of images have been tampered with maliciously, which seriously affects the authenticity of image content. Different from the general object detection, the study of image tampering detection needs to pay more attention to the tamper information of the image itself, which is often manifested in a weak form. Therefore, the detection model needs to focus on learning more abundant tamper features. This paper proposes a dual-stream Faster R-CNN model that combines gradient edge information and attention mechanism, and the model can realize detection and location of regions with different tampering types. One of the two streams is the color stream, which uses the attention mechanism to extract the surface features of the image, such as brightness contrast, visual difference of tampering with the boundary, etc. The second of the two streams is a gradient stream. A Gradient high-pass filter is used to enhance the anomaly edge features between the real area and the tampered area, making it easier for the model to find faint tampered traces in the tampered image. Finally, the features of color stream and gradient stream are fused by means of compact bilinear pooling. Due to the relatively small size of publicly available image tampering data sets, the Pascal VOC 2012 is used to create an image tampering detection data set which containing 10 010 images for model pre-training. The experimental results on COVER, Columbia, and CASIA data sets show that the model proposed in this paper improves the detection accuracy by 7.1% to 9.6% compared to the latest models, and exhibits higher robustness under JPEG compression and image blur attacks. [ABSTRACT FROM AUTHOR]

Local feature information plays an important role in image segmentation, however, the referring image segmentation task is dependent on the text expression, so it is impossible to extract local feature information directly from the original reference image. In order to solve this problem, this paper proposes a specific noun-guided local feature extraction deep neural network model（NgLFNet). The NgLFNet model can automatically mine the local feature information of the object to be segmented according to the key nouns in the input text expression. Specifically, the model first obtains key nouns in text through sentence analysis; secondly, extracting corresponding features through text and image encoders, and using the key nouns to obtain local features of high-interest regions through the multihead attention mechanism; then the multi-modal features are gradually fused to learn; Finally, the decoding and correction module uses the obtained local features to perform more detailed corrections on the prediction mask to obtain the final result. The method in this paper is compared with a variety of mainstream referring segmentation methods and the experimental results show that the proposed method improves the accuracy of text-based instance segmentation task. [ABSTRACT FROM AUTHOR]

With the explosive growth of video data, the task of video captioning had been paid more and more attention by researchers. How to enable computers to understand the content of the video and express it accurately was one of the difficult problems that had not been solved perfectly in the field of video captioning task. Aiming at the problems of insufficient use of semantic information and inaccurate description in the video captioning model, a video captioning method combining semantic enhancement and multi-attention mechanism was proposed based on the encoder-decoder framework in this paper. Firstly, visual text feature aggregation was used to provide high-level semantic guidance for model coding. Then, the Faster-RCNN network was used to extract the features of the video object, and the potential semantic information of the video object was obtained through the graph convolutional network, resulting in enhanced features. Finally, a multiple attention mechanism was introduced to better utilize input information and enhance the learning ability of the model. The experimental results on MSVD and MSR- VTT data sets showed that, compared with the benchmark model, the proposed method can reasonably optimize the input information of the video description model, effectively extract the video potential semantics, and solve the cross-modal problem of video text, and resolve the syntax structure of generated statements. This method can effectively improve the accuracy of video description model and the ability to describe complex scenes, showing its advancement. [ABSTRACT FROM AUTHOR]

In the realm of intensive aquaculture, real-time monitoring of fish feeding status is one of the important bases for developing scientific feeding strategies. To achieve temporal detection of fish feeding states, and to solve the problem of high error in detecting feeding state switching by existing methods, this paper proposes a fish state detection network BMN-Fish based on an improved BMN (Boundary Matching Network), which introduces an efficient depth module and consists of an Efficient Channel Attention Module Net (ECANet) and a Base Residual Module (BRM). Among them,the Efficient Channel Attention Module can expand the perceptual field of the temporal dimensional features,to improve the ability of the model to extract local information; the Base Residual Module enables the algorithm to focus on the regions of interest in the feature map, which can strengthen the global perception ability of the algorithm. The experimental results showed that the AUC and AR@ 100 of BMN-Fish were 93. 32% and 95. 28%,which were 2. 17% and 1. 95% higher than BMN,respectively. BMN-Fish could effectively detect the sequential feeding movements of fish, and provide important information for formulating intelligent feeding strategies for intensive aquaculture. [ABSTRACT FROM AUTHOR]

Complementary product recommendation aims to provide users with products that are often purchased together to meet common needs. Existing complementary product recommendation methods mainly consider modeling the content characteristics（visual and textual content） of products, but do not consider the preferences of users when purchasing products. To this end, this paper designed a complementary product recommendation model that integrates user preferences, termed as CPRUP（complementary product recommendation models that integrate user preferences）.The model first computed the complementarity of image and text features between products; then it combined knowledge graph with attention mechanism, mines the correlation between user’s historical purchased products based on n-hop neighbors, and proposed a user representation based on knowledge graph to extract user preferences for complementary products; finally, it implemented joint learning of complementarity and user preferences based on neural networks. Experiments on the Amazon dataset show that the proposed CPRUP model has increased the ACC by 5%and the Precision by 4% compared to the suboptimal baseline model, indicating that the CPRUP model can more accurately recommend complementary products to users. [ABSTRACT FROM AUTHOR]

In the field of music information retrieval (MIR), classification according to music genres is a challenging task. Traditional audio feature engineering methods requires manually selecting and extracting music signal features for processing, resulting in complex feature extraction process, unstable model performance and poor generalization. The method combining deep learning with spectrogram also has some problems such as unsuitable model for some data and difficulty in global feature extraction. This paper proposes a music genre classification model based on convolutional attention mechanism, called MGTN. MGTN combines two music genre classification methods: input spectrogram and audio signal feature extraction, to construct audio time series data, which greatly improves the model's ability to extract features and generalization, and provides a new idea for music genre classification. Experimental results on GTZAN and Ballroom datasets show that the MGTN model can effectively fuse input data from two different modalities. Compared with dozens of benchmark models, the MGTN model has strong advantages. [ABSTRACT FROM AUTHOR]

In the driving scenario, due to the occlusion between pedestrians and their scale variations, detection model have low accuracy, high model parameters, and difficulty in deploying to mobile terminals. This paper proposes a lightweight real-time pedestrian detection model, LPD-YOLO, based on the YOLOv5s model. Firstly, in the feature extraction part, the original backbone network is replaced with MES Net (Mish-Enhanced Shuffle Net), and an attention module SA (Shuffle Attention) is embedded in the backbone network to enhance network feature extraction ability. Secondly, in the feature fusion part, the original PANet is improved by using the DS-ASFF structure to fully fuse feature maps of different sizes. Then, standard convolution is replaced with GS convolution in the feature-covergent network part without affecting accuracy, further reducing model parameters and computation. Finally, in the prediction part, the original loss function is improved by using the OTA label assignment strategy combined with α-IOU to accelerate model convergence. Experimental data show that compared with YOLOv5s, LPD-YOLO has 81.2% fewer parameters, 46.3% lower floating-point operation volume, 75.8% smaller model size, and 3.3% higher detection accuracy. The single image detection speed is 13.2 ms, which better meets the real-time detection requirements of dense pedestrians in driving scenarios. [ABSTRACT FROM AUTHOR]

The paper introduces a refined dual attention mechanism designed to mitigate attention deficiency challenges encountered during the segmentation of diminutive objects within the domain of autonomous driving. In this context, multifaceted factors such as lighting, weather, and road conditions create intricate challenges. The primary objective of this approach is to augment the capacity for feature representation by incorporating position and channel attention mechanisms to derive weights. Initially, the position attention mechanism discerns the significance of each pixel within the spatial domain, generating position-related weights. Subsequently, the channel attention mechanism assesses the importance of each channel in feature representation, resulting in channel-related weights. These derived position and channel attention weights are multiplicatively applied to the input features on an element-wise basis, thereby enhancing their representation capabilities. The resultant features from the two attention modules are amalgamated. Experimental findings substantiate that this enhanced network model significantly elevates the accuracy of semantic segmentation. It achieves an average intersection over union (mIoU) of 80.4% on the Cityscapes dataset, marking a substantial improvement of 10.4% in comparison to the baseline fully convolutional network(FCN)method. [ABSTRACT FROM AUTHOR]

Aerial image has the characteristics of large amount of data, small target scale and dense distribution, and its angle of view is downcast, which is different from the head up view of ordinary image. Therefore, the traditional target detection algorithm for ordinary image cannot adapt to the target detection task of aerial image. For small target detection in aerial images, an improved algorithm based on YOLO-v5, Small-Tiny-YOLO-v5, was proposed. Firstly, GhostNet network was used as the backbone network of the improved algorithm; secondly, the attention mechanism module was added in the backbone network; thirdly, an aerial image data set for small targets was constructed. In addition, the idea of transfer learning was integrated in the training of the improved algorithm. Experimental results show that the model parameters of the proposed improved algorithm are lower than the original YOLO algorithm,and the accuracy and speed are also better than the original algorithm.In the public dataset and the dataset constructed in this paper,the accuracy has increased by 0.009and0.024respectively,and the speed has increased by 73.735% and 58.641%respectively. [ABSTRACT FROM AUTHOR]

Shield load is the main performance indicator of the shield, accurate load prediction is significant to ensure the safety and efficiency of the shield and the stability of the surrounding environment. Recognizing the limitations of the traditional prediction methods, this paper proposes a hybrid model (CBM), combining convolutional neural network (CNN), bi-directional long short-term memory (BiLSTM) and attention mechanism, to predict the shield load accurately based on the high-dimensional feature and time series characteristic of the data. The proposed model not only can extract the high-dimensional features and time series characteristics of the data, but also can highlight the importance of high-dimensional features and important time node information. The experiment results show that compared with the existing models, the proposed model achieves a higher prediction performance, the prediction accuracy of the thrust and torque is 94.2% and 96.2%. [ABSTRACT FROM AUTHOR]

For the beamforming design of intelligent reflector assisted large-scale MISO systems, most of the current researches are basing on the fully known channel state information, and do not consider the coupling relationship between the active beamforming matrix at the base station and the reflection coefficient matrix of intelligent reflector. The intelligent reflector assisted channel is a cascade channel, which is difficult to estimate and has high pilot frequency overhead. In view of this, this paper proposed a hybrid intelligent reflective surface architecture containing active and passive components. Firstly, based on the sparse characteristics of millimeter-wave channels, it used the compressed sensing algorithm to estimate the channels, and designed a two-stage convolutional network framework integrating attention mechanism to jointly optimize the reflection coefficient matrix and beamforming matrix with the aim of maximizing and rate. The experimental results show that compared to existing methods, the proposed method can achieve better sum rate performance with fewer pilots, effectively reducing pilot loss and reducing computational time complexity. And when the communication environment changes, the network also has good robustness. [ABSTRACT FROM AUTHOR]

To address the issues of traditional video summarization algorithms not fully utilizing multimodal information in videos and struggling to ensure temporal consistency of summary video segments, this paper proposed a new video summarization algorithm(MTNet) that fused multimodal features and temporal zone detection. Firstly, it extracted the visual and audio features of the videos using pre-trained GoogLeNet and VGGish models, and designed a dimension smoothing operation to align the two modal features, endowing the model with comprehensive representation capabilities. Secondly, considering that the generated video summaries should have global representativeness, it combined single and double-layer self-attention mechanisms with residual structures. to extract long-range temporal features of video images and audio features, obtaining a single vector representation of the model in the temporal domain. Lastly, it predicted summary boundaries and importance of indivi- dual temporal segments in the video using separated temporal zone detection and weight sharing methods. It selected key video segments to generate video summaries through non-maximum suppression. Experimental results show that MTNet has stronger representation capabilities and robustness on two standard datasets, SumMe and TvSum. It achieves an increase in F1-score compared to the anchor-free video summarization algorithm DSNet-AF and the shot importance prediction-based video summarization algorithm VASNet on both datasets. [ABSTRACT FROM AUTHOR]

To address the problem of unbalanced computational resource consumption and accuracy of semantic segmentation models using attention mechanism, this paper proposed a lightweight attention enhancement algorithm for semantic segmentation. Firstly, it designed a striped dimensional attention mechanism based on the shape characteristics of objects in driving scenes, used striped pooling instead of traditional square convolution, and combined dimensionality reduction operations to extract long-range semantic associations in each dimension to cut down the model computation. Then it fused the attention on channel domain and spatial domain to form a lightweight multidimensional attention fusion module that could be superimposed and disassembled to extract feature information in all directions and further improve the model accuracy. Finally, it inserted the module into the ResNet-101 backbone based encoding-decoding network to guide the semantic fusion of high and low layers, correct the feature map edge information, and supplement the prediction details. The experiments show that the module has strong robustness and generalization ability, cutting about 90% of the number of parameters and 80% of the computation compared with the same type of attention mechanism, and the segmentation accuracy still achieves a stable improvement. [ABSTRACT FROM AUTHOR]

To address the semantic consistency problem of migrated code in the source code migration model, this paper introduced the statement-level attention mechanism based on the token-level attention mechanism, proposed a source code migration model HPGN based on the hierarchical attention mechanism, and designed state feeding mechanisms. During migration, the statement-level attention mechanism aligned the features of source code statements and migrated code statements, the token-level attention mechanism extracted tokens from the aligned code statements, and the state feeding mechanism passed the feature of adjacent migrated code statement, thus improving the semantic consistency of migrated code. Experimenting on a real project dataset, the results show that HPGN improves the overall score by 3.4 over the best comparison model while having fewer model parameters. In addition, ablation experiments validate the effectiveness of the state feeding mechanisms and HPGN hierarchical architecture. [ABSTRACT FROM AUTHOR]

Aspect-level sentiment classification aims to discriminate the sentiment polarity of a specific aspect in a sentence. Although attention-based recurrent neural network models perform well among existing solutions, they are not ideal for processing "semantically ambiguous" sentences that are short and contain many neologisms and polysemous words. Therefore, this paper proposes a neural network model based on knowledge graph and attention mechanism. The basic idea is to use a knowledge base to obtain a relevant concept set of aspect words and integrate external information to enhance the semantic representation of the text. Firstly, the output of bidirectional long short-term memory network is combined with self-attention mechanism to generate context representation. Then, the upper and lower context representations are combined to use dual attention to obtain external knowledge from the knowledge graph and obtain knowledge vectors related to aspect words. Finally, the two parts of content are input together into a fully connected network to calculate the aspect-level sentiment tendency. Experimental results show that compared with other models, the proposed model significantly improves classification performance. [ABSTRACT FROM AUTHOR]

In recent years, the deep learning network model based on UNet and GAN (generative adversarial network) has shown unique advantages in the field of image restoration. But, there are still artifacts, blurring, degradation of texture details, difficult to repair large-area damage, and the repaired holes are incompatible with the background image. In order to solve the problems that the current existing models are not friendly to the repair of large-area damaged images and the degraded images after repair, this paper improved the Shift-UNet model by studying the existing methods. On the basis of U-Net and GAN, this paper added an improved attention mechanism Attention-UNet between each layer of encoder and decoder and integrated it into the Shift-UNet to form Attention-Shift-UNet. And through research, it changed the activation function of the original down-sampling part from Leaky＿ ReLU to SiLU. The improved model not only achieved good results on the 64×64 central cover, but also realized the random cover, and the cover area increased from 20 % to 80 %. The experimental results show that the repair effect of the proposed model is better, especially for the repair of large-area damaged images. After testing on CelebA, Paris Architecture and Paris Stree View datasets, the evaluation indexes have been significantly improved, and SSIM has increased from 0.944 5 to 0.947 1. PSNR is increased from 27.992 7 to 28.553 6. L2 loss is decreased from 0.001 7 to 0.001 5. [ABSTRACT FROM AUTHOR]

Multimodal fusion algorithm is one of the solutions to the problem of urban road video caption which only considers visual information and ignores the equally important audio information. Existing multimodal fusion algorithms based on Transformer all have the problem of low fusion performance between modes and high computational complexity. In order to improve the interaction between multimodal information, This paper recently proposed a new Transformer based model called Multimodal Attention Bottleneck for Video Captioning (MABVC). Firstly, this paper uses pre-trained I3 D and VGGish networks to extract visual and audio features of video and input the extracted features into Transformer model. Then, the decoder part will train the information of the two modes respectively and perform multimodal fusion. Finally, the model processes the results of the decoder and generates text captions that people can understand. This paper conducted a comparison experiments using data sets MSR-VTT, MSVD and self-built data sets BUUISE, and validated model results using evaluation metrics the model. The experimental results show that the video caption model based on multimodal attention fusion has obvious improvement in all indicators. The model can still achieve good results on traffic scene data sets, and has great application prospects/can be promoted and applied in intelligent driving industry. [ABSTRACT FROM AUTHOR]

In order to further mine the changelog information, this paper proposed a siamese Bi-lSTM network model based on the attention mechanism to classify the changelog to realize topic annotation and assist in locating the location of code defects. The model proposed a word segmentation tool with security features to realize changelog preprocessing, used Bi-lSTM network to learn changelog contextual semantic information, and solved the problem of overfilling mode existing in changelog itself through siamese neural network and expanded the data set with high quality to improve generalization ability. It carried out serialization training for the changelog composed of multiple sentences, and the influence of sentences was distinguished through the attention mechanism. For some changelogs of the defect repair class, this paper improved the LLVM tool, generated a mapping table to search the log content, and located the location of the defect module in the source code. A large number of experimental results show that the classification effect of the model in this paper has strong generalization ability, and is nearly 10% higher than the general methods in text classification methods in terms of accuracy, F, value and other indicators, has ideal log classification effect and topic learning effect. [ABSTRACT FROM AUTHOR]

To address the problem of poor recommendation accuracy in the session recommendation model, this paper proposed a capsule-based graph convolutional disentanglement for session-aware recommendation (CGCD) method.Specifically, it used disentangled learning technique to transition item embedding into factor embedding based on multiple sub-channels. Then, it used a capsule dynamic fusion strategy to aggregate different factors to obtain a new item embedding.In addition, it used a multi-head attention mechanism to assign weights to each item in the session. Finally, it aggregated the item embedding with other items in the current session according to the assigned weights, and then generated an accurate session representation to achieve item recommendation. Experiments on two publicly available real datasets show that the proposed model can improve the Pre@10, Pre@20, MRR@10 and MRR@20 by 5.17%, 2.99%, 6.56% and 2.94% on average of the recommended performance, which verifies the effectiveness and efficiency of the method in this paper. [ABSTRACT FROM AUTHOR]

YOLO-based mussel （Mytilus edulis） recognition and detection technology is the key to achieving mechanization and intelligence in the grading, seedling separation, and other operational links of mussels. However, the unclear and indefinite external features of mussels present a challenge to improving recognition accuracy. This paper proposes a mussel object detection model based on the improved YOLOv5 algorithm （CST-YOLO）, which integrates the CoordAttention attention mechanism to enhance feature expression ability, uses the SIoU boundary box regression Loss function to reduce the boundary box regression loss and improve the detection speed of the model, and designs an improved decoupled head TSCODE to enhance the detection accuracy. The algorithm testing was performed on a self-constructed dataset of mussels, and the experimental results demonstrated that the CST-YOLO algorithm exhibited an improvement in accuracy （P） by 0.428% compared to the YOLOv5 algorithm, and mAP （mean Average Precision） at IoU thresholds of 0.5 and 0.95 reached 92.221%, indicating a significant increase of 1.583%. These findings highlight the effective enhancement of mussel object detection accuracy while ensuring optimal detection speed achieved by the CST-YOLO algorithm. The present study contributes to the advancement of machine vision technology in facilitating the automation and intelligent production and processing within the mussel farming industry. [ABSTRACT FROM AUTHOR]

Coarse-grained attention mechanisms based on global semantic interactions cannot effectively use the interconnection between modalities to extract the key parts of modal information, which affects classification results.Aiming at this problem, this paper proposes a modal information interaction model, which extracts the local semantic association features of the modality through a fine-grained attention mechanism and uses them for sentiment classification. First, the intra-modal information interaction module constructs intra-modal connections and generates intra-modal interaction features. After this, the inter-modal information interaction module uses intra-modal information interaction features to generate gating vectors, which focus on the appropriate parts of the text（image）, thereby obtaining inter-modal information interaction. Finally, considering the redundant information in the features, the model also adds an adaptive feature fusion module to select features from the global feature level, enhance the expression ability of key features containing emotional information, and weaken the impact of redundant information on classification results. Experimental results on two public datasets MVSA-Single and MVSA-Multi demonstrate that the model outperforms a series of baseline models. [ABSTRACT FROM AUTHOR]

Existing lip-reading algorithms for Chinese sentences are inadequate at feature extraction and visual ambiguity resolution, which leads to low accuracy. Aiming at this problem, this paper proposed a lip-reading algorithm for Chinese sentences based on temporal convolutional network and 3D convolutional neural network（3DT-CHLipNet）.Firstly, it used a Temporal Convolutional Network to extract long-term features from lip dynamics sequences, which has a much larger receptive field than the Long Short Term Memory Network. Secondly, in order to minimize the visual ambiguity in Chinese lipreading, it adopted a Transformer model with the self-attention mechanism to capture the context information and improve the accuracy of sentence prediction. Finally, it introduced a temporal masking data augmentation strategy in the data preprocessing to further reduce the error rate of the algorithm. Comparison experiments on CMLR, the largest open-source Sequence-to-Sequence Chinese Mandarin Lip Reading dataset, show that the improvement in accuracy over representative lip reading algorithms for Chinese sentences ranges from 2.17%to 23.99%. [ABSTRACT FROM AUTHOR]

Objectives: Sea-land segmentation is of great significance for tasks such as ocean target detection and coastline extraction in synthetic aperture radar (SAR) image. To solve the problem of sea-land segmentation of multi-resolution SAR image in practical applications, this paper presents a sea-land segmentation method based on context and edge attention. Methods: The proposed method uses the channel attention mechanism to fuse context features of different scales and levels, and designs an edge extraction branch to provide edge information for further improving the segmentation result of the boundary area. In addition, a sea-land segmentation dataset of multi-resolution SAR image based on the Gaofen-3 satellite data is provided. The dataset covers multiple resolution images, including various sea-land boundary types such as ports, islands. Using this multi-resolution SAR image coastline segmentation dataset, we perform experimental analyses on the effectiveness of the proposed network and the contributions of each module. Results and Conclusions: Experimental results show that the proposed method can work well for the task of sealand segmentation, the overall classification accuracy and mean intersection over union achieve 98.21% and 96.47%, respectively. [ABSTRACT FROM AUTHOR]

Aiming at the problem of large number of parameters and low detection efficiency of the traditional target detection method in defect detection of electronic components, this paper proposes a target detection method based on the lightweight YOLOX detection network. Firstly, the backbone network is lightened using deeply separable convolution to reduce parameters and improve detection speed. Secondly, a spatial Pyramid-based channel attention model is constructed to filter and fuse features of different scales to enhance the feature weights of small size defects. In the feature fusion upsampling process, efficient channel attention is added to improve detection accuracy with slightly increased parameters. Finally, the EIoU loss function is used to optimize the IoU loss function, and the cosine annealing algorithm is used to make the model achieve the best detection effect. The model is tested on a self-made dataset of electronic component appearance defects, and the average detection accuracy reaches 98.96%, with a detection time of approximately 0.09 seconds per image. Compared with the original model, the detection speed is doubled and the model size is reduced by about 60%. The model is also validated on the PCB defect public dataset, achieving fast detection of target defects. [ABSTRACT FROM AUTHOR]

Air pollution seriously affects human health and social sustainable development. However, the multi-variable air quality data obtained by sensors often have missing values, which brings difficulties to data analysis and processing. Currently, many analysis methods for changes in a certain air component only rely on time data and spatial data of this attribute, ignoring the influence of other air components on the trend of this attribute in the same time interval. In addition, it is difficult to achieve ideal results in filling discrete missing data. This paper proposes a Time Attention Model (TAM) based on deep learning, which uses attention mechanism to focus on the correlation between different timestamps and the correlation between different feature time series, and combines short-term historical data to fill missing values in multi-variable air quality data. The proposed model is evaluated using air quality data from Beijing, and the experimental results show that TAM has advantages over ten other baseline models. [ABSTRACT FROM AUTHOR]

Aiming at the problems of inaccurate prediction of normal video and poor ability of learning normal features in video anomaly detection, this paper proposed an anomaly detection model combining wavelet transform and encoder-decoder attention. The model introduced multi-level discrete wavelet transform, and designed a module of discrete wavelet transform fusion. The module concatenated the sub-bands obtained by decomposing video frames, and fed the result into depthwise separable convolution, and then fused with the encoder features to compensate for the high-frequency details lost in the down sampling process. The model also constructed encoder-decoder attention module. After performing difference of Gaussian operation on the encoder feature map, the attention weights were obtained along the horizontal and vertical directions respectively. And then the encoder features were aggregated according to the weights. Finally, the decoder features were associated to enhance the network’s learning of normal events. Experiments on Ped1,Ped2 and Avenue datasets show that the AUC of the proposed model is increased by 3.2%,3.1% and 2.0%.And the results indicate that the proposed model can effectively improve the abnormal detection ability. [ABSTRACT FROM AUTHOR]

To realize intelligent cancer diagnosis for pancreatic images returned by endoscopic ultrasonography, the classic classification and recognition network AlexNet and SE attention mechanism are combined with in this paper. The proposed SE-AlexNet network can accurately detect cancerous images and discriminate the pancreatic parts to which the normal images belong. In order to verify the reliability and superiority of the algorithm, a series of comparison experiments are carried out under the self-made dataset in the paper, including the model comparison experiments between the basic network AlexNet and other classical classification networks, and the comparison experiments of various models improved and obtained by inserting different attention mechanism in different positions. The results show that the overall precision and recall rate of SE-AlexNet model can reach 99.56% and 98.69%. For cancerous images, the detection precision and recall rate are 100%, which can achieve an effective auxiliary diagnosis for practical endoscopic ultrasonography of pancreatic cancer. [ABSTRACT FROM AUTHOR]

Most of the existing image tampering detection methods are only for a certain tampering method, and there is a problem that the detection accuracy of the tampered area boundary is not high. In this regard, this paper proposed a U-shaped network based on dual-stream encoder-decoder architecture for image tampering. Firstly, the method used the skip connection between the encoder and the decoder to fuse the low-level and high-level features in the tampered image, and used the atrous convolution and CBAM attention mechanism to fuse the features output by the encoder, so that the network had better localization performance for tampered regions of different scales; Secondly, in order to improve the network’s detection accuracy of the boundary of the tampered area, this algorithm used the image morphological method to make a tampered boundary dataset; Finally, it used multiple loss functions to optimize the performance of the network simultaneously, namely it used cross-entropy and root-mean-square loss functions to measure the tampered region loss and tampered boundary loss of the prediction graph respectively. The experimental results on 4 public datasets: CASIA, Columbia, NIST16, Coverage show that the method proposed in this paper can effectively detect the tampering area of the fake image with splicing and copy-paste tampering methods, and output pixel-level tampering region location map. Compared with other mainstream tampering detection methods, the proposed method achieves the highest AUC values on the CASIA and Columbia dataset and the highest F1 values on the Columbia dataset. [ABSTRACT FROM AUTHOR]