Search

Paper recommendation with personalized methods helps researchers to track the latest academic trends and master cutting-edge academic trends efficiently. Meanwhile, the methods of previous paper recommendation suffer from three problems: data sparsity of content-based and collaborative filtering methods; Graph-based recommendations do not fully consider the personalized information of authors and their articles; Cold start based on deep learning. To overcome those difficulties, we propose a personalized paper recommendation method based on a knowledge graph and Transformer encoder (KGTE) with a self-attention mechanism. Firstly, we add auxiliary information (article title, publication year, citation times, and abstract) as attributes to the nodes of knowledge graph(KG), which contain author, digital object unique identifier(DOI) and keywords. Secondly, BERT is used to represent the semantic information features of the article and Transformer is introduced to fully integrate the feature context. After that, by using RippleNet, we traverse the knowledge graph, filter the user preference distribution and form a set of pre recommended nodes with multi_hop nodes. Finally, the prediction layer sorts the set and gets a Top_n paper recommendation. In the experiments on the DBLP and Aminer datasets, the precision value of KGTE improved by an average of 2.59% over the existing baseline methods DER and 4.23% improvement in NDCG. [ABSTRACT FROM AUTHOR]

During the past decades, several academic paper recommendation systems were introduced in literature aiming to assist users in finding relevant papers close to their needs. In particular, those models have used users' past preferences combined with papers' side information to personalize their recommendations. Unfortunately, the majority of those models fail to utilize the implicit relationship between the crucial elements along with the semantic relations of the nodes in a Heterogeneous Information Network (HIN), which can further improve the accuracy of the models. In this paper, we propose a research article recommendation model that aims to tackle those issues by exploiting both textual and graph representations, simultaneously. The model employs SPECTER document embedding to learn context-preserving research article representations. In particular, it learns the features from a HIN, which is a knowledge graph of the node entities and their relationships. Then, an attention module strengthens the semantic feature extraction process and learns enhanced representations. The combined semantic and structural features are then provided as input to a Deep Neural Network (DNN) in order to learn high-level representations of query and candidate papers. We have evaluated our model against state-of-the-art models over two popular datasets. The results indicate a significant improvement in terms of MAP, recall, and MRR. [ABSTRACT FROM AUTHOR]

The process of retrieving suitable papers which are related to an interesting topic while doing research normally takes a lot of time and effort. Citation recommendation is frequently used to solve this problem by automatically suggesting a list of candidate papers that should match with the user's references or topics of interest. Applying the recent research results of deep learning in multiple disciplines, the performance of citation recommendation systems has been significantly improved with the facilitation of powerful deep neural data analysis and representation learning techniques. However, most of the recent Neural Citation Network (NCN) model-based techniques still encounter limitations related to the capability of integrating auxiliary information to assist the citation contextual learning process. Thus, in this paper, we propose a novel context-aware NCN-based model with the extra textual data integration and Bidirectional Encoder Representations from Transformers (BERT) model to improve the performance of the citation recommendation task. To do this, an extensive deep neural auto-encoding mechanism with a self-attention-based mechanism is utilized in our proposed model to flexibly learn both associated textual and citation contextual data in the given dataset. These enriched citation contextual information representations are then utilized to improve the performance of the citation recommendation task as the end-to-end neural learning process. Extensive experiments in the standard arXiv dataset show the effectiveness and good performance of the proposed model. [ABSTRACT FROM AUTHOR]

With the development of image feature matching technology, feature matching algorithms based on deep learning have achieved excellent results, but in scenarios with low texture or extreme perspective changes, the matching accuracy is still difficult to guarantee. In this paper, a superresolution reconstruction method based on a Residual-ESPCN (efficient subpixel convolutional neural network) approach is proposed based on LoFTR (local feature matching with transformers). The superresolution method is used to improve the interpolation method used in ASFF (adaptive spatial feature fusion) to increase the image resolution, enhance the detailed information of the image, and make the extracted features richer. Then, ASFF is introduced into the local feature extraction module of LoFTR, which can alleviate the inconsistency problem of information transmission between different scale features of the feature pyramid and lessen the amount of information lost during transmission from low- to high-resolution levels. Moreover, to improve the adaptability of the algorithm to different scenarios, OTSU is introduced to adaptively calculate the threshold of feature matching. The experimental results show that in different indoor or outdoor scenarios, our proposed algorithm for matching features can effectively improve the adaptability of feature matching and can achieve good results in terms of the area under the curve (AUC), accuracy and recall. [ABSTRACT FROM AUTHOR]

In the field of food safety, milk has become an indispensable beverage in people's lives. Therefore, it is of great significance to detect milk freshness. Starting from the milk freshness detection, the perspective can be extended to liquid detection. Liquid detection has attracted much attention and has been applied in many fields. However, current liquid detection methods are either contact detection methods that can lead to sample damage, or require specialized instruments, expertise for operation or cumbersome hardware deployment. This paper introduces a non-contact milk freshness detection method based on image features and tensor construction. Unlike existing liquid detection methods, our method relies on ubiquitous WiFi signals to achieve non-contact and non-invasive milk freshness detection. The design intuition is that the WiFi signals will lead to different multipath propagation when passing through milk with different freshness, which can be used to detect milk freshness. We use existing commercial devices to collect WiFi signal data of milk with different freshness, denoise the collected data and transform the denoised data into multiple types of time-frequency images and spatial-temporal images, and input the images into the deep learning network to extract image features that contain richer and more comprehensive information, and then utilize the extracted image features to perform tensor construction to better retain the original time, frequency and spatial feature information, and apply 3D convolutional layer and fully connected layers for milk freshness detection. The experimental results show that our method achieves a high accuracy of 93.25% in detecting milk freshness. [ABSTRACT FROM AUTHOR]

Schizophrenia (SZ) is a prevalent mental disorder characterized by cognitive, emotional, and behavioral changes. Symptoms of SZ include hallucinations, illusions, delusions, lack of motivation, and difficulties in concentration. While the exact causes of SZ remain unproven, factors such as brain injuries, stress, and psychotropic drugs have been implicated in its development. SZ can be classified into different types, including paranoid, disorganized, catatonic, undifferentiated, and residual. Diagnosing SZ involves employing various tools, including clinical interviews, physical examinations, psychological evaluations, the Diagnostic and Statistical Manual of Mental Disorders (DSM), and neuroimaging techniques. Electroencephalography (EEG) recording is a significant functional neuroimaging modality that provides valuable insights into brain function during SZ. However, EEG signal analysis poses challenges for neurologists and scientists due to the presence of artifacts, long-term recordings, and the utilization of multiple channels. To address these challenges, researchers have introduced artificial intelligence (AI) techniques, encompassing conventional machine learning (ML) and deep learning (DL) methods, to aid in SZ diagnosis. This study reviews papers focused on SZ diagnosis utilizing EEG signals and AI methods. The introduction section provides a comprehensive explanation of SZ diagnosis methods and intervention techniques. Subsequently, review papers in this field are discussed, followed by an introduction to the AI methods employed for SZ diagnosis and a summary of relevant papers presented in tabular form. Additionally, this study reports on the most significant challenges encountered in SZ diagnosis, as identified through a review of papers in this field. Future directions to overcome these challenges are also addressed. The discussion section examines the specific details of each paper, culminating in the presentation of conclusions and findings. [ABSTRACT FROM AUTHOR]

Furnace temperature and oxygen content are important parameters reflecting the combustion inside a circulating fluidized bed (CFB) boiler. Accurately predicting boiler output is a complex task due to the high noise and nonsmoothness of actual boiler input and output data. In this paper, a new hybrid convolutional neural network (CNN), bidirectional long short-term memory (biLSTM) network, and squeezing and excitation (SE) network prediction model is proposed to significantly improve the prediction accuracy of oxygen content and furnace temperature by combining the advantages of multiple deep learning networks. This network can extract spatiotemporal characteristics of input parameters such as coal feed to effectively predict boiler furnace temperature and oxygen content. CNNs can extract complex features such as dynamic and static nonlinearities between multiple variables affecting the furnace temperature and oxygen content, as well as high noise. The biLSTM network layer can efficiently handle the temporal information of irregular trends in modeling time series components; SE can extract the important information between channels through the feature relationships between channels for better overfeature extraction. The CNN-biLSTM-SE model can effectively solve the problem of nonlinear mapping complexity between inputs and outputs. Experiments show that the proposed CNN-biLSTM-SE model outperforms existing methods. The experimental results showed that the average MAPE errors for oxygen content prediction were CNN-biLSTM-SE (0.038), CNN-biLSTM with attention mechanism (AM) (0.043), CNN-biLSTM (0.051), CNN-LSTM (0.051), biLSTM (0.051), RNN (0.051), LSTM(0.0052), and CNN(0.0054). Extensive experiments in CFB boilers with oxygen content and furnace temperature show that the proposed CNN-biLSTM-SE model achieves better results in terms of goodness-of-fit, generalization ability and accuracy. [ABSTRACT FROM AUTHOR]

Marine mammals are an important part of marine ecosystems, and human intervention seriously threatens their living environments. Few studies exist on the marine mammal call recognition task, and the accuracy of current research needs to improve. In this paper, a novel MG-ResFormer two-channel fusion network architecture is proposed, which can extract local features and global timing information from sound signals almost perfectly. Second, in the input stage of the model, we propose an improved acoustic feature energy fingerprint, which is different from the traditional single feature approach. This feature also contains frequency, energy, time sequence and other speech information and has a strong identity. Additionally, to achieve more reliable accuracy in the multiclass call recognition task, we propose a multigranular joint layer to capture the family and genus relationships between classes. In the experimental section, the proposed method is compared with the existing feature extraction methods and recognition methods. In addition, this paper also compares with the latest research, and the proposed method is the most advanced algorithm thus far. Ultimately, our proposed method achieves an accuracy of 99.39% in the marine mammal call recognition task. [ABSTRACT FROM AUTHOR]

The authors suggest to trade-off diagnosis quality versus runtime performance, exploiting learning for constraint ordering and where matrix factorization is used to estimate diagnoses. The final manuscript, "Applying matrix factorization to consistency-based direct diagnosis", authored by Seda Polat Erdeniz et al., deals with diagnosis as a basis for configuration problems when efficient solving is required in real-world applications. Screening all the accepted papers, the top 7 authors of the accepted papers were encouraged to submit their papers for this special issue, asking for substantial extensions to their submissions for creating manuscripts that are appropriate for being published in Applied Intelligence. [Extracted from the article]

As malicious code attacks continue to evolve, attackers leverage techniques like packing and code obfuscation to generate numerous variants, challenging traditional detection methods. Addressing the limitations of current deep learning-based malicious code classification approaches in feature extraction and accuracy, this paper introduces an innovative RGB visualization detection method based on a hybrid multi-head attention mechanism. Initially, a feature representation method utilizing RGB images is introduced. This approach focuses on semantic relationships between a malware's binary information, assembly details, and API data, generating images with richer textural information. This technique effectively uncovers the deep dependencies between the original and variant versions of malicious code, providing stronger support for subsequent classification tasks. Furthermore, to tackle the issues of malware encryption and obfuscation, a deep neural network framework is adopted, incorporating a modular design philosophy and integrating a multi-head attention mechanism. This design not only enhances the expressiveness of critical features but also helps the model better focus on key aspects of the malicious code, thereby improving classification accuracy. Through comparative experiments and in-depth analysis, the effectiveness and superiority of the proposed RGB visualization method and MSA-ResNet model in the field of malicious code variant classification are validated. The accuracy rates achieved on the Kaggle and DataCon datasets are 99.49% and 97.70%, respectively, representing significant improvements over other methods. This approach demonstrates strong generalization capabilities and resistance to obfuscation, offering a new and effective tool for malicious code detection. [ABSTRACT FROM AUTHOR]

In the cloud computing domain, significant strides have been made in performance prediction for cloud workflows, yet link prediction for cloud workflows remains largely unexplored. This paper introduces a novel challenge: joint node and link prediction in cloud workflows, with the aim of increasing the efficiency and overall performance of cloud computing resources. GNN-based methods have gained traction in handling graph-related tasks. The unique format of the DAG presents an underexplored area for GNNs effectiveness. To enhance comprehension of intricate graph structures and interrelationships, this paper introduces two novel models under the DAGCN framework: DAG-ConvGCN and DAG-AttGCN. The former synergizes the local receptive fields of the CNN with the global interpretive power of the GCN, whereas the latter integrates an attention mechanism to dynamically weigh the significance of node adjacencies. Through rigorous experimentation on a meticulously crafted joint node and link prediction task utilizing the Cluster-trace-v2018 dataset, both DAG-ConvGCN and DAG-AttGCN demonstrate superior performance over a spectrum of established machine learning and deep learning benchmarks. Moreover, the application of similarity measures such as the propagation kernel and the innovative GRBF kernel-which merges the graphlet kernel with the radial basis function kernel to accentuate graph topology and node features-reinforces the superiority of DAGCN models over graph-level prediction accuracy conventional baselines. This paper offers a fresh vantage point for advancing predictive methodologies within graph theory. [ABSTRACT FROM AUTHOR]

Non-invasive electroencephalography (EEG) is a technique for monitoring brain activity that is valuable in the diagnosis and study of the brain. However, due to factors such as brain-computer interface (BCI) devices deficiency, dynamic network limitations, and subject issues, EEG signals may fade throughout the entire process from signal generation to acquisition. The fading of EEG signals can cause changes in the data distribution, blur the information and have a negative impact on subsequent research and application. In order to reduce the adverse effects of data fading, this paper proposes a hierarchical bidirectional long-short term memory (LSTM)-Attention network based on cognitive brain limbic system (HBLANet). HBLANet classifies randomly fading signals multiple times by way of a dimensionality-reducing classification algorithm in order to narrow down the feature interval processed by a single neural network. Then the different types of EEG signals acquired from the classification are processed in a more focused manner using a bidirectional LSTM-Attention network. In this paper, the overall performance of the network is greatly improved by decomposing complex signal processing tasks into smaller tasks. The model is evaluated on the EEG-denoisenet dataset and compared with competitive networks such as feature pyramid network (FPN), UNet, MultiResUNet, 1D-ResCNN etc., and the results show that the proposed network achieves better fading processing outcomes. In the overall experiment, the processed EEG signals achieve the relative root mean squared error (RRMSE) value of 0.009, the signal-to-noise ratio (SNR) of 32.78, and the correlation coefficient (CC) of 0.98. Furthermore, the denoising task in the overall experiment achieves even more exciting results, with a SNR of 40.48 and a CC of 0.991 for the processed EEG signal in the case of processing high SNR signals (the range of the SNR is from -2 to 2). Therefore, we believe that the framework has an important reference value for future research on signal quality restoration. [ABSTRACT FROM AUTHOR]

Camouflage object detection (COD) is designed to locate objects that are "seamlessly" embedded in the surrounding environment. Camouflaged object detection is a challenging task due to the high intrinsic similarities between objects and their backgrounds, as well as the low boundary contrast between them. To address this problem, this paper proposes a new ternary symmetric fusion network (TSFNet), which can detect camouflaged objects by fully fusing features of different levels and scales. Specifically, the network proposed in this paper mainly contains two key modules: the location-attention search (LAS) module and the ternary symmetric interaction fusion (TSIF) module. The location-attention search module makes full use of contextual information to position potential target objects from a global perspective while enhancing feature representation and guiding feature fusion. The ternary symmetric interaction fusion module consists of three branches: bilateral branches gather rich contextual information of multi-level features, and a middle branch provides fusion attention coefficients for the other two branches. The strategy can effectively achieve information fusion between low- and high-level features, and then achieve the refinement of edge details. Experimental results show that the method is an effective COD model and outperforms existing models. Compared with the existing model SINetV2, TSFNet significantly improves the performance by 3.5% weighted F-measure and 8.1% MAE on the COD10K. [ABSTRACT FROM AUTHOR]

Deep learning has made substantial progress in crowd counting, but in practical applications, due to interference factors such as perspective distortion and complex background, the existing methods still have large errors in counting. In response to the above problems, this paper designs a multi-scale feature fusion network (IA-MFFCN) based on the reverse attention mechanism, which maps the image to the crowd density map for counting. The network consists of three parts: feature extraction module, inverse attention module, and back-end module. First, to overcome the problem of perspective distortion, deeper single-column CNNs was designed as a feature extraction module to extract multi-scale feature information and merge them; second, to avoid interference of complex backgrounds, the inverse attention module was designed, through the multi-scale inverse attention mechanism, reducing the influence of noise on counting accuracy. Finally, to generate a high-quality crowd density map, dilation convolution was introduced. Simultaneously, to enhance the sensitivity of the network to crowd counting, a comprehensive loss function based on Euclidean loss and predicted population loss is designed to improve training accuracy, to produce a more accurate density value. Experiments show that compared with the comparison algorithm, the algorithm in this paper has a significant reduction in the mean absolute error (MAE) and mean square error (MSE) on the ShanghaiTech dataset, UCF_CC_50 dataset and WorldExpo'10 dataset. [ABSTRACT FROM AUTHOR]

In recent years, the Multi-agent Deep Reinforcement Learning Algorithm has been developing rapidly, in which the value method-based algorithm plays an important role (such as Monotonic Value Function Factorisation (QMIX) and Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement learning (QTRAN)). In spite of the fact, the performance of current value-based multi-agent algorithm under complex scene still can be further improved. In value function-based model, a mixing network is usually used to mix the local action value of each agent to get joint action value when the partial observability will cause the problem of misalignment and unsatisfying mixing results. This paper proposes a multi-agent model called Transform Networks that transform the individual local action-value function gotten by agent network to individual global action-value function, which will avoid the problem of misalignment caused by partial observability when the individual action value is mixed, and the joint action value can represent the cooperative conditions of all agents well. Using the StarCraft Multi-Agent Challenge (SMAC) as the experimental platform, the comparison of the performance of algorithms on five different maps proved that the proposed method has better effect than the current most advanced baseline algorithms. [ABSTRACT FROM AUTHOR]

Dunhuang murals are a collection of Chinese style and national style, forming an autonomous Chinese-style Buddhist art. It has a very great historical and cultural value and an important research value. Among them, the lines of Dunhuang murals are extremely general and expressive. It reflects the character's distinct character and intricate inner emotions. Hence, the outline drawing of the murals is of great importance in search of the Dunhuang culture. The generation of contours of Dunhuang murals belongs to the detection of image contours, which is an important branch of computer vision, aimed at extract salient contour information from the images. Although convolution-based deep learning networks have achieved good results in extracting image contours by exploring the contextual and semantic features of images. However, as the receptive field broadens, some detailed local information is lost. As a result, it is impossible for them to generate reasonable outline drawings of murals. In this paper, we propose a novel edge detector based on self-attention combined with convolution to generate line drawings of Dunhuang murals. Compared with existing edge detection methods, firstly, a new residual self-attention and convolution mixed module(Ramix) is proposed to merge local and global features into feature maps. Secondly, a novel densely connected backbone extraction network is designed to efficiently propagate information rich in edge features from shallow layers into deep layers. It is apparent from the experimental results that the method of this paper can generate richer and sharper edge maps on several standard edge detection datasets. In addition, tests on the Dunhuang mural dataset show that our method can achieve very competitive performance. [ABSTRACT FROM AUTHOR]

Deep metric learning (DML) models aim to learn semantically meaningful representations in which similar samples are pulled together and dissimilar samples are pushed apart. However, the classification effect is limited due to the high time complexity of previous models and their poor performance in extracting data relationships. This paper presents a novel relationship constraint deep metric learning (RCDML) approach, including proxy relationship constraint (PRC) and sample relationship constraint (SRC) for inter-class separability and intra-class compactness, to solve the above problems and improve the classification effect. The PRC combines the proxy-to-proxy relationship loss term with the proxy-to-sample relationship loss function to maximize the proxy features, hence enhancing inter-class separability by decreasing proxy similarity. Additionally, the SRC combines the sample-to-sample relationship loss term with the proxy-to-sample relationship loss function to maximize the sample features, which promotes intra-class compactness by increasing the similarity between the most different samples of the same class. Unlike existing proxy-based and pair-based methods, the relationship constraint framework uses a diverse range of proxy and sample data relationships. In addition, the proxy correction (PC) module is used to optimize the proxy. Extensive tests conducted on the widely popular CUB-200-2011, CARS-196, and SOP datasets show that the framework is effective and attains state-of-the-art performance. [ABSTRACT FROM AUTHOR]

The fusion of multimodal medical images has garnered painstaking attention for clinical diagnosis and surgical planning. Although various scholars have designed numerous fusion methods, the challenges of extracting substantial features without introducing noise and non-uniform contrast hindered the overall quality of fused photos. This paper presents a multimodal medical image fusion (MMIF) using a novel deep convolutional neural network (D-CNN) along with preprocessing schemes to circumvent the mentioned issues. A non-linear average median filtering (NL-AMF) and multiscale improved top-hat (MI-TH) approach are utilized at the preprocessing stage to remove noise and improve the contrast of images. The non-linear anisotropic diffusion (NL-AD) scheme is employed to split the photos into base and detailed parts. The fusion of base parts is accomplished by a dimension reduction method to retain the energy information. In contrast, the detailed parts are fused by novel D-CNN to preserve the enriched detailed features effectively. The simulation results demonstrate that the proposed method produces better brightness contrast and more image details than existing methods by acquiring 0.7649 to 0.8986, 0.3520 to 0.4783, 0.7639 to 0.9056, 68.8932 to 81.0487 gain for quality transfer ratio from source photo to a generated photo ( Q G AB ), feature mutual information (FMI), structural similarity index (SSIM), and average pixel intensity (API) respectively. [ABSTRACT FROM AUTHOR]

In unstructured scenarios, objects usually have unique shapes, poses, and other uncertainties, which put forward higher requirements for the robot's planar grasping detection ability. Most previous methods use single-modal data or simply fused multi-modal data to predict gripper configurations. Single-modal data is not conducive to comprehensively describe the diversity of objects, and the simple fusion method may also ignore the dependencies between multi-modal data. Based on the above considerations, we propose a Multi-modal Dynamic Cooperative Fusion Network (MDCNet), in which a Multilevel Semantic Guided Fusion Module (MSG) is designed, through which enhanced semantic guidance vectors are used to suppress the undesired influence factors produced by different fusion structures. In addition, we also design a general Enhanced Feature Pyramid Nets Structure (EFPN) to learn the dependencies between fine-grained features and coarse-grained features and improve the robustness of the encoder in unstructured scenarios. The results show that the proposed method has an accuracy rate of 98.9% on the Jacquard dataset and 99.6% on the Cornell dataset. In over 2000 robotic grasp trials, our structure achieves a grasp success rate of 98.8% in single-object scenarios and 93.5% in cluttered scenarios. The proposed method in this paper is superior to previous grasp detection methods in both speed and accuracy, and has strong real-time performance. [ABSTRACT FROM AUTHOR]

Few-shot learning is a critical task in computer vision processing that helps reduce deep learning's reliance on large datasets. This paper aims to establish a few-shot learning network that is adaptive to diverse scenes. A novel approach referred to as task-adapted network with bi-directional long short-term memory network (BiLSTM-TANet) is proposed in this paper. BiLSTM-TANet is an end-to-end approach based on deep metric learning and designed to use the information from finite samples as much as possible. It fuses the context embeddings and structure information of the images and adaptively adjusts the features several times during the feature extraction of task to achieve task-specific embedding and quickly adapt to different distributed tasks, improves the feature extraction performance, and strikes a balance between model stability and generality. The model employs Euclidean distance as the classifier to reduce the number of model parameters and enhance the classification performance. Experiments conducted on miniImageNet, TieredImageNet, CUB200_2011 and CIFAR-FS datasets demonstrate the performance of the proposed BiLSTM-TANet. Furthermore, the effects of different few-shot learning parameters on the model's performance are explored, providing a helpful reference for the future study of few-shot learning. Finally, a series of ablation studies are performed to analyze the performance of BiLSTM-TANet. [ABSTRACT FROM AUTHOR]

Few-shot learning is a critical task in computer vision processing that helps reduce deep learning's reliance on large datasets. This paper aims to establish a few-shot learning network that is adaptive to diverse scenes. A novel approach referred to as task-adapted network with bi-directional long short-term memory network (BiLSTM-TANet) is proposed in this paper. BiLSTM-TANet is an end-to-end approach based on deep metric learning and designed to use the information from finite samples as much as possible. It fuses the context embeddings and structure information of the images and adaptively adjusts the features several times during the feature extraction of task to achieve task-specific embedding and quickly adapt to different distributed tasks, improves the feature extraction performance, and strikes a balance between model stability and generality. The model employs Euclidean distance as the classifier to reduce the number of model parameters and enhance the classification performance. Experiments conducted on miniImageNet, TieredImageNet, CUB200_2011 and CIFAR-FS datasets demonstrate the performance of the proposed BiLSTM-TANet. Furthermore, the effects of different few-shot learning parameters on the model's performance are explored, providing a helpful reference for the future study of few-shot learning. Finally, a series of ablation studies are performed to analyze the performance of BiLSTM-TANet. [ABSTRACT FROM AUTHOR]

In unstructured scenarios, objects usually have unique shapes, poses, and other uncertainties, which put forward higher requirements for the robot's planar grasping detection ability. Most previous methods use single-modal data or simply fused multi-modal data to predict gripper configurations. Single-modal data is not conducive to comprehensively describe the diversity of objects, and the simple fusion method may also ignore the dependencies between multi-modal data. Based on the above considerations, we propose a Multi-modal Dynamic Cooperative Fusion Network (MDCNet), in which a Multilevel Semantic Guided Fusion Module (MSG) is designed, through which enhanced semantic guidance vectors are used to suppress the undesired influence factors produced by different fusion structures. In addition, we also design a general Enhanced Feature Pyramid Nets Structure (EFPN) to learn the dependencies between fine-grained features and coarse-grained features and improve the robustness of the encoder in unstructured scenarios. The results show that the proposed method has an accuracy rate of 98.9% on the Jacquard dataset and 99.6% on the Cornell dataset. In over 2000 robotic grasp trials, our structure achieves a grasp success rate of 98.8% in single-object scenarios and 93.5% in cluttered scenarios. The proposed method in this paper is superior to previous grasp detection methods in both speed and accuracy, and has strong real-time performance. [ABSTRACT FROM AUTHOR]

The automation of hand hygiene monitoring is critical in healthcare for ensuring clean hands and preventing infectious disease spread. While advancements have been made, existing methods have limitations in accurately detecting and classifying handwashing actions. This paper addresses these limitations and introduces the Residual-Based Multilevel Fused Network (ResMFuse-Net) as a novel approach to automate the quality assurance of hand hygiene procedures. Our model integrates advanced techniques, including feature fusion, model compression, a feature fusion block (FFB), and a modified separable residual block (SE-ResB). The proposed model fused two networks into one trainable feature extraction pipeline, and applies model compression to retain the core blocks that are crucial for propagating strong and robust features while conserving a significant fraction of the computing resources. Additionally, we introduce a FFB that includes ConvLSTM and alpha dropout to learn spatial dependencies, establish correlations between frames in a video, and mitigate overfitting. This paper introduces a SE-ResB, which is a customized residual component composed of separable convolutions and LeakyReLU activation. The SE-ResB is incorporated to handle the fused features and generate a more diverse set of features, leading to considerable performance enhancements. This study also includes an ablation analysis that highlights the importance of each component. The proposed ResMFuse-Net is evaluated on two datasets: a newly created handwashing dataset (451 videos) and a publicly available dataset (656 videos). Achieving a recognition accuracy of 97.61% on the handwashing dataset and 98.69% on the other dataset, the ResMFuse-Net outperforms previous methods with fewer parameters and FLOPs, demonstrating its efficiency and cost-effectiveness. [ABSTRACT FROM AUTHOR]

Rigid point cloud registration is a crucial upstream task in computer vision, whose goal is to align two misaligned point clouds using a rigid transformation. Existing methods, which directly utilize extracted point features for computing point relationships, are likely to result in a wrong-matching relationship since two or more similar feature points in the source point cloud easily correspond to the same point in the target point cloud. To this end, this paper proposes a two-view point cloud registration network that better alleviates the problem of similar feature points from both the feature and geometry levels. Specifically, at the feature level, a residual correction unit is proposed to learn feature-aware coefficients from raw 3D point clouds to adaptively increase or decrease the difference between features. An attention mechanism is established in two-point clouds to capture the implicit feature relationship between the two-point clouds, gathering the information of another point cloud to enrich the feature information of its own point cloud. Second, at the geometry level, a dual-view graph topology fusion module is described. All points in the graph structure are no longer independent but connected by their geometry structure. Therefore, each point can aggregate neighbor information in the same point cloud and in different point clouds through the constructed single graph and interactive graph, so that each point can enhance the difference between points through its geometry structure. In order to fuse the information of a single graph and an interactive graph, a cross-attention module is proposed to supplement contextual information to obtain point features that are more suitable for matching. Experimental results demonstrate that our method achieves excellent results on complete and partial noisy point clouds. [ABSTRACT FROM AUTHOR]

Currently, generative applications are reshaping different fields, such as art, computer vision, speech processing, and natural language. The computer science personalization area is increasingly relevant since large companies such as Spotify, Netflix, TripAdvisor, Amazon, and Google use recommender systems. Then, it is rational to expect that generative learning will increasingly be used to improve current recommender systems. In this paper, a method is proposed to generate synthetic recommender system datasets that can be used to test the recommendation performance and accuracy of a company on different simulated scenarios, such as large increases in their dataset sizes, number of users, or number of items. Specifically, an improvement in the state-of-the-art method is proposed by applying the Wasserstein concept to the generative adversarial network for recommender systems (GANRS) seminal method to generate synthetic datasets. The results show that our proposed method reduces the mode collapse, increases the sizes of the synthetic datasets, improves their ratings distributions, and maintains the potential to choose the desired number of users, number of items, and starting size of the dataset. Both the baseline GANRS and the proposed Wasserstein-based WGANRS deep learning architectures generate fake profiles from dense, short, and continuous embeddings in the latent space instead of the sparse, large, and discrete raw samples that previous GAN models used as a source. To enable reproducibility, the Python and Keras codes are provided in open repositories along with the synthetic datasets generated to test the proposed architecture (https://github.com/jesusbobadilla/ganrs.git). [ABSTRACT FROM AUTHOR]

Trajectory prediction for moving objects is a critical task for intelligent transportation with numerous applications, such as route planning, traffic management, congestion alleviation, etc. In this paper, we propose a novel framework that integrates sequence modeling, trajectory clustering and topology extraction to improve the accuracy of trajectory prediction. By incorporating self-attention for sequence modeling, we are able to effectively capture the temporal dependencies in trajectory data. Additionally, by taking into account the clustering information via a variational auto-encoder and the topological information based on a graphical neural network (GNN), we can further improve the accuracy of trajectory prediction. Furthermore, integrating a GNN facilitates our framework to handle diverse characteristics of road networks, such as road distance and traffic status, thereby making the proposed approach adaptive to different practical scenarios. As demonstrated by the experimental results on two publicly available datasets, our proposed method improves the accuracies by up to 0.5% and 3.8% for 1-step and 15-step predictions respectively, compared to the state-of-the-art method. [ABSTRACT FROM AUTHOR]

Time series data, characterized by large volumes and wide-ranging applications, requires accurate predictions of future values based on historical data. Recent advancements in deep learning models, particularly in the field of time series forecasting, have shown promising results by leveraging neural networks to capture complex patterns and dependencies. However, existing models often overlook the influence of short-term cyclical patterns in the time series, leading to a lag in capturing changes and accurately tracking fluctuations in forecast data. To overcome this limitation, this paper introduces a new method that utilizes an interpolation technique to create a fine-scaled representation of the cyclical pattern, thereby alleviating the impact of the irregularity in the cyclical component and hence enhancing prediction accuracy. The proposed method is presented along with evaluation metrics and loss functions suitable for time series forecasting. Experiment results on benchmark datasets demonstrate the effectiveness of the proposed approach in effectively capturing cyclical patterns and improving prediction accuracy. [ABSTRACT FROM AUTHOR]

Co-training is a semi-supervised algorithm that aims to improve prediction effects by exchanging confident instances and pseudo-labels among multiple learners. One central issue is how to mitigate the negative impact of low-quality pseudo-labels during training. In this paper, we propose semi-supervised regression via embedding space mapping and pseudo-label smearing (S2RMS) to ensure that unlabeled data contribute positively to the prediction process. First, a Triplet neural network is trained using pairwise data generated from labeled data. This network maps the training data to the embedding space to better separate dissimilar instances. Second, the embedded data are randomly partitioned into different subsets to train corresponding regression models (a.k.a. regressors). These regressors are integrated into the prediction process. Third, these subsets are augmented using unlabeled data with high similarity to the labeled data and high-confidence pseudo-labels. Here, the similarity and confidence are calculated using the network and the smearing technique, respectively. Experiments are conducted on fourteen datasets, and the results are compared to those of three excellent algorithms. The results show that S2RMS outperforms other co-training and metric semi-supervised regression algorithms. The source code is available at: https://github.com/BetaCatPro/S2RMS. [ABSTRACT FROM AUTHOR]

The design of effective multi-view clustering algorithms has recently garnered significant research attention. In this paper, we develop a robust multi-view clustering via collaborative constraints and multi-layer concept factorization (RMCCMCF) model to enhance clustering power and robustness by mining low-rank data and hierarchical information from multiple views. This approach adopts low rank (LR) and deep learning (DL) to explore the essential and hierarchical structures hidden in various views. The proposed approach applies multiple graph regularization constraints to extract manifold information from different views. Kernel technology is used for concept factorization in high-dimensional space to effectively distinguish different classes of sample points. Additionally, we design an alternating iterative optimization approach to solve the RMCCMCF algorithm and discuss its convergence. The results of extensive experiments conducted on the CBSR, HW2Sources, 3Sources, CUB, and UCI datasets demonstrate that the RMCCMCF approach outperforms several other recent multi-view clustering approaches in terms of recognition performance. [ABSTRACT FROM AUTHOR]

Breast cancer remains a substantial public health challenge, marked by a rising prevalence. Accurate early detection is paramount for effective treatment and improved patient outcomes in breast cancer. The diversity of breast tumors and the complexity of their microenvironment present significant challenges. Establishing a reliable breast calcification and micro-calcification detection approach is an ongoing issue that researchers must continue to investigate. The goal is to develop an effective methodology that contributes to increased patient survival. Therefore, this paper introduces a novel approach for classifying breast calcifications in mammography, aiming to distinguish between benign and malignant cases. Aiming to address these challenges, we proposed our hybrid approach for breast calcification classification in mammogram images. The proposed approach starts with an image pre-processing phase that includes noise reduction and enhancement filters. Afterward, we proposed our hybrid classification architecture. It includes two branches: First, the vision transformer (ViT++) branch for contextual features. Secondly, a CNN branch based on transfer learning techniques for visual features. Using the CBIS-DDSM dataset, the application of our proposed ViT++ architecture reached the maximum accuracy of 96.12%. Further, the application of the VGG16 as a single feature extractor had a much lower accuracy of 61.96%. Meanwhile, the combination of these techniques in the same architecture improved the accuracy to 99.22%. Three different pre-trained feature extractors were applied in the CNN branch: Xception, VGG16, and RegNetX002. However, the best-obtained outcomes were from the combination of the ViT++ and the VGG16. The experimental findings indicate that the proposed strategy for breast calcification detection has the potential to surpass the performance of currently top-ranked methods, particularly in terms of classification accuracy. [ABSTRACT FROM AUTHOR]

Deep learning technology has been successfully applied in more and more fields. In this paper, the application of deep neural networks in higher-order nonparametric spatial autoregressive models is studied. For spatial model, we propose the higher-order nonparametric spatial autoregressive neural network (HNSARNN) to fit the model. This method offers both good interpretability and prediction performance, and solves the black box problem in deep learning models to some degree. In various scenarios of spatial data distribution, the proposed method demonstrates superior performance compared to traditional approaches for handling nonparametric functions (such as the B-spline method). Simulation results show the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

The windspeed on the sea surface is an important factor affecting the process of detecting sea surface targets, as it affects the reflection and radiation intensity, posing challenges for precise detection of optical remote sensing targets. This study collects and obtains ultra-high-resolution hyperspectral data using a push-broom hyperspectral sensor, and employs a novel data processing approach to process the raw data and extract significant features related to sea surface windspeed. Based on this foundation, further exploration is conducted into the practical requirements of the spatial dimension and spectral resolution of remote sensing sensors. Dimensionality reduction processing is applied to the hyperspectral data in both spatial and spectral dimensions. It is noteworthy that certain low-resolution spectral extractions will be derived from the central wavelengths of AVIRIS and Landsat satellites, and generate spatial-spectral joint heatmaps (SSHMs) capable of depicting sea surface windspeeds, so as to assess the potential of windspeed estimation utilizing these sensors. Moreover, four kinds of convolutional neural network (CNN) architectures, in conjunction with transfer learning techniques, are utilized to achieve precise estimation of sea surface windspeed. The final estimation results show that CNN models achieve reliable performance until the spectral dimension is reduced to 8 or below, and the spatial dimension is reduced to 25; The lightweight framework effectively employs coarse-resolution SSHMs to extract windspeed and achieve accurate recognition results. This paper introduces a feasibility study for windspeed estimation utilizing the optical sensors that meet the spatial and spectral resolution requirements. [ABSTRACT FROM AUTHOR]

In this paper, we proposed a novel deep image coding and decoding model for GMMQI (Gaussian mixture model quartile intervals, GMMQI) and a variable rate bit allocation method to optimize the rate distortion performance and prioritize the transmissions of more significant information. First, to address the problem of bit streams leading to ambiguity during encoding and decoding, we convert the image into a potential tensor, each element of which uses a four-bit parameter dictionary to preserve the parameter bits. Then, the variable rate is calculated using quaternions based on the parameter dictionary, and a hybrid CLM &BAM (Channel latent map & Bit allocation map) approach is designed to assign bits to the potential tensor and encode it, which transforms the problem of finding the optimal encoder-decoder into finding the optimal hyper-parameters in the model and reduces the complexity of the GMMQI model. Finally, a GMMQI approach with variable rate bit allocation is developed in combination with CLM &BAM, to be able to prioritize the transmission of more significant information. The experimental results show that the GMMQI method reaches an advanced level compared to the traditional image compression standards BPG, JPEG2000, and JPEG, and is comparable to the most advanced level compared to the existing deep learning based compression methods. [ABSTRACT FROM AUTHOR]

In the dynamic smart grid landscape, accurate probabilistic forecasting of electric load is critical. This paper presents a novel 24-hour-ahead probabilistic load forecasting model by integrating quantile regression with a parallel convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) architecture. Carefully tuning hyperparameters can enhance model performance and generalization capability. Consequently, we propose an improved whale optimization algorithm for automatic hyperparameter tuning of the forecasting model. Case studies demonstrate the proposed method's superior performance over benchmark models in terms of average interval score and pinball loss. In addition, it exhibits valid coverage and tight interval bandwidths. The model provides precise short-term load forecasts to support robust smart grid planning and operations. [ABSTRACT FROM AUTHOR]

3D target detection is a research hotspot in recent years. In the field of autonomous driving, 3D target detection is mainly targeted at outdoor scenes that the camera height is constant. In a few indoor scenes, 3D target detection is mostly at the category level. However, it is difficult to generate instance-level 3D target detection datasets. In complex indoor scenes, instance-level 3D target detection is used as the research object in this paper. The indoor 3D target detection dataset is constructed by Aruco marker. A pixel-by-pixel key point voting network for joint semantic segmentation of RGB images is established, and a new key point assumption strategy is proposed. Combined with depth images, the key point detection is extended to three dimensions, and the bit pose is optimized by the ICP algorithm. The evaluation metrics and visualization of the model are analyzed and compared. It is tested and visualized under validation set, truncated validation set and unlabeled. The generalization of the method in this paper is proved, and 3D target detection in indoor scene based on RGB image and RGB-D image is achieved. [ABSTRACT FROM AUTHOR]

The purpose of active learning is to significantly reduce the cost of annotation while ensuring the good performance of the model. In this paper, we propose a novel active learning method based on the combination of pool and synthesis named dual generative adversarial active learning (DGAAL), which includes the functions of image generation and representation learning. This method includes two groups of generative adversarial network composed of a generator and two discriminators. One group is used for representation learning, and then this paper performs sampling based on the predicted value of the discriminator. The other group is used for image generation. The purpose is to generate samples which are similar to those obtained from sampling, so that samples with rich information can be fully utilized. In the sampling process, the two groups of network cooperate with each other to enable the generated samples to participate in sampling process, and to enable the discriminator for sampling to co-evolve. Thus, in the later stage of sampling, the problem of insufficient information for selecting samples based on the pool method is alleviated. In this paper, DGAAL is evaluated extensively on three data sets, and the results show that DGAAL not only has certain advantages over the existing methods in terms of model performance but can also further reduces the annotation cost. [ABSTRACT FROM AUTHOR]

This paper presents a comprehensive review of deep learning applications in the video game industry, focusing on how these techniques can be utilized in game development, experience, and operation. As relying on computation techniques, the game world can be viewed as an integration of various complex data. This examines the use of deep learning in processing various types of game data. The paper classifies the game data into asset data, interaction data, and player data, according to their utilization in game development, experience, and operation, respectively. Specifically, this paper discusses deep learning applications in generating asset data such as object images, 3D scenes, avatar models, and facial animations; enhancing interaction data through improved text-based conversations and decision-making behaviors; and analyzing player data for cheat detection and match-making purposes. Although this review may not cover all existing applications of deep learning, it aims to provide a thorough presentation of the current state of deep learning in the gaming industry and its potential to revolutionize game production by reducing costs and improving the overall player experience. [ABSTRACT FROM AUTHOR]

Human ideas and sentiments are mirrored in facial expressions. Facial expression recognition (FER) is a crucial type of visual data that can be utilized to deduce a person's emotional state. It gives the spectator a plethora of social cues, such as the viewer's focus of attention, emotion, motivation, and intention. It's said to be a powerful instrument for silent communication. AI-based facial recognition systems can be deployed at different areas like bus stations, railway stations, airports, or stadiums to help security forces identify potential threats. There has been a lot of research done in this area. But, it lacks a detailed review of the literature that highlights and analyses the previous work in FER (including work on compound emotion and micro-expressions), and a comparative analysis of different models applied to available datasets, further identifying aligned future directions. So, this paper includes a comprehensive overview of different models that can be used in the field of FER and a comparative study of the traditional methods based on hand-crafted feature extraction and deep learning methods in terms of their advantages and disadvantages which distinguishes our work from existing review studies.This paper also brings you to an eye on the analysis of different FER systems, the performance of different models on available datasets, evaluation of the classification performance of traditional and deep learning algorithms in the context of facial emotion recognition which reveals a good understanding of the classifier's characteristics. Along with the proposed models, this study describes the commonly used datasets showing the year-wise performance achieved by state-of-the-art methods which lacks in the existing manuscripts. At last, the authors itemize recognized research gaps and challenges encountered by researchers which can be considered in future research work. [ABSTRACT FROM AUTHOR]

As an essential part of geocoding, address matching has gained increasing research attention. Due to the long-distance dependency and unstructured property, existing address-matching methods hardly capture the contextual and implicit semantic information of unstructured text addresses. This paper presents a Contrastive Learning model based on Global and Local representation Interaction for Address Matching (referred to as CLGLIAM), which introduces a novel global and local interaction network to enhance the discrimination ability of the model on the hard negative address by associating the relationship between the global and local address representation explicitly. Simultaneously, to improve the generalization and transferability of the model, we utilize contrastive learning to enrich the data sample and extricate the model from task-specific knowledge. Furthermore, extensive experiments are conducted on Shenzhen and national address datasets to verify the effectiveness of our approach. Our model achieves state-of-the-art F1 scores of 99.26 and 98.50 on the two datasets, respectively. And the extended hard negative experiments further demonstrate the better performance of CLGLIAM in terms of semantic discrimination. [ABSTRACT FROM AUTHOR]

In order to solve the problem that traditional two-way decision based malicious code detection methods fail to consider the influence of decision making under the condition of insufficient information when facing complex and massive data in dynamic environment, this paper proposes a malicious code detection model based on sequential three-way decision. This model introduces sequential three-way decision into the domain of malicious code to avoid the risk of possible error detection due to insufficient information. In order to improve the overall performance of the detection model and eliminate the subjectivity of threshold selection, this paper designs a multi-objective sequential three-way decision model based on the above model, while considering the decision efficiency and decision accuracy of the model. In addition, the multi-objective optimization algorithm is used to solve the model. The simulation results show that the model not only guarantees the detection performance, but also improves the decision efficiency effectively. The real dynamic detection environment is better fitted. [ABSTRACT FROM AUTHOR]

Intelligent monitoring of machine chips can help the machine tool to remove chip accumulation in time. Based on chip shape information, real-time monitoring of the machine tool is also possible. Although image segmentation algorithms have been widely used in industry, they still need to improve. These methods still have some limitations in effectively combining global and local information of industrial images, and there needs to be an image segmentation algorithm specifically for machine tool chips. In this paper, we propose a deep learning-based algorithm for lightweight chip image segmentation named Hybridformer. First, the algorithm generates scale-aware semantic features using feature maps of different scales as the input part. Then, we use the cross specification and SelfNorm to improve the generalization robustness of the model in terms of distribution bias. Finally, the network extracts global contextual information fusing EfficientNetv2 features to achieve accurate localization. Widespread experimental results show that Hybridformer outperforms both CNN and Transformer algorithms in terms of comprehensive performance on multiple chip image datasets, and it can achieve a good balance between accuracy and size of parameters. [ABSTRACT FROM AUTHOR]

For humans to trust in artificial intelligence (AI) systems, it is essential for machine learning (ML) models to be interpretable to users. For example, the judicial process requires that AI conclusions must be rigorous and absolutely interpretable. In this paper, we propose a novel approach, VAE-SLIME, for providing stable local interpretable model-agnostic explanations (SLIME) based on a variational autoencoder (VAE). LIME is a technique that explains the predictions of any classifier in an interpretable and faithful manner. Despite the great success of LIME, the most popular method in this category, it has several disadvantages due to its random perturbation-based sampling method. The VAE-SLIME proposed in this paper is specifically designed to address the lack of stability and local fidelity exhibited by LIME for tabular data. VAE-SLIME first employs fixed noise to replace the random Gaussian noise used by the reparameterization trick of the VAE. Then, it uses this new VAE model instead of random perturbation method to generate stable samples. By considering the sequential relationship and flipping of features, a novel explanation stability evaluation metric, the feature sequence stability index (FSSI), is introduced to accurately evaluate the stability of explanations. In a comparison with 6 state-of-the-art approaches on 7 commonly used tabular datasets, the experimental results show beyond doubt that the explanations produced by our approach are most stable, and its local fidelity is 65.17% higher than that of other approaches on average. [ABSTRACT FROM AUTHOR]

Improve the scientific level of the goose breeding industry and help the development of intelligent agriculture. Instance Segmentation has a pivotal role when the breeders make decisions about geese breeding. It can be used for disease prevention, body size estimation and behavioural prediction, etc. However, instance segmentation requires high performance computing devices to run smoothly due to its rich output. To ameliorate this problem, this paper constructs a novel encoder-decoder module and proposes the SDSCNet model. The reasonable use of depth-separable convolution in the module reduces the number and size of model parameters and increase execution speed. Finally, SDSCNet model enables real-time identification and segmentation of individual geese with the accuracy reached 0.933.We compare this model with numerous mainstream instance segmentation models, and the final results demonstrate the excellent performance of our model.Furthermore, deploying SDSCNet model on the embedded device Raspberry Pi 4 Model B can achieve effective detection of continuous moving scenes. [ABSTRACT FROM AUTHOR]

As an emerging biometric recognition technology, gait recognition has the advantages of non-contact long distance and difficult to imitate. Existing gait recognition methods perform gait recognition by using features extracted from the overall appearance or local regions of humans. However, the detailed features extracted by current gait recognition methods based on human local region lose the overall relevance of the image and the edge information of human local region. Secondly, the method based on the local area of the human body does not focus on the local parts of the human body that are less affected by clothing occlusion. To solve the above problems, this paper proposes a new gait recognition network framework GaitLRDF, which improves the accuracy and robustness of gait recognition by Local Relation Convolutional layers (LRConv) and Human Body Focusing module(HBF). LRConv can simultaneously use the global and local information of the human body, and the local detail features extracted in the module can retain the edge information of the human body. HBF can focuse on the gait parts that are less affected by clothing occlusion, and obtain more discriminative gait detail features. The experimental results show that in the three gait environments of NM, BG and CL set by CASIA-B dataset, GaitLRDF is 0.40%, 0.10% and 1.10% higher than the current most advanced method respectively. The recognition accuracy on OU-MVLP dataset reaches 91.40%. [ABSTRACT FROM AUTHOR]

Structural deep document clustering methods, which leverage both structural information and inherent data properties to learn document representations using deep neural networks for clustering, have recently garnered increased research interest. However, the structural information used in these methods is usually static and remains unchanged during the clustering process. This can negatively impact the clustering results if the initial structural information is inaccurate or noisy. In this paper, we present an adaptive structural enhanced representation learning network for document clustering. This network can adjust the structural information with the help of clustering partitions and consists of two components: an adaptive structure learner, which automatically evaluates and adjusts structural information at both the document and term levels to facilitate the learning of more effective structural information, and a structural enhanced representation learning network. The latter incorporates integrates this adjusted structural information to enhance text document representations while reducing noise, thereby improving the clustering results. The iterative process between clustering results and the adaptive structural enhanced representation learning network promotes mutual optimization, progressively enhancing model performance. Extensive experiments on various text document datasets demonstrate that the proposed method outperforms several state-of-the-art methods. The overall framework of adaptive structural enhanced representation learning network [ABSTRACT FROM AUTHOR]

The development of a new energy generation industry, represented by wind power generation, constitutes a crucial endeavor in attaining the dual carbon goal. Nonetheless, the high-dimensional, nonlinear, and stochastic nature of wind speed sequences imposes substantial demands on power systems' peak regulation capability. To address the limitations of traditional prediction methods, which struggle to provide probabilistic prediction outcomes and exhibit poor generalizability, this paper introduces a fusion of fuzzy theory and advanced artificial intelligence algorithms. The proposed approach, termed the fuzzification-forecasting-optimization ensemble system (FFOES), enables both point and interval predictions of wind speed, thereby complementing the emerging literature on power system dispatch. By leveraging fuzzy set theory and the concept of information granulation, the original sequence is initially divided into manageable particles of smaller scales. Subsequently, the computational efficiency of shallow neural networks and the automatic feature extraction capability of deep learning are harnessed to construct a comprehensive and integrated strategy for wind speed prediction. The numerical results demonstrate that the FFOES achieves a superior balance between prediction accuracy and stability compared to that of 15 benchmark models. Notably, the system's robust performance across distinct sites and time intervals underscores the significant potential of the FFOES in the realm of wind system risk management. [ABSTRACT FROM AUTHOR]

As the use of drones continues to increase, their capabilities pose a threat to airspace safety when they are misused. Deploying AI models for intercepting these unwanted drones becomes crucial. However, these AI models, such as deep learning models, often operate as "black boxes", making it hard to trust their decision-making system. This also affects end-users' confidence in these AI systems. In this paper, the explainability of deep reinforcement learning is investigated and a deep reinforcement learning (DRL) method, double deep Q-network with dueling architecture and prioritized experience replay is applied to train the AI models. To make the AI model decisions more transparent and to understand the reasoning behind the AI decisions for counter-drone systems, Shapley Additive Explanations (SHAP) method is implemented. After training the DRL agent, experience replay is visualized, and the absolute SHAP values are calculated to explain the key factors that influence the deep reinforcement learning agent's choices. The integration of DRL with explainable AI methods such as SHAP demonstrates significant potential for the advancement of robust and efficient counter-drone systems. [ABSTRACT FROM AUTHOR]

Multi-sensor information fusion (MSIF) technology based on deep convolutional neural networks (CNN) has been widely used in UAV obstacle avoidance. However, detection efficiency needs to be improved in practice because of its high computational complexity and limited airborne hardware resources. A lightweight CNN-based fast detection method based on the fusion of millimeter-wave (MMW) radar and camera is proposed in this paper. In the data preprocessing stage, the input images to the network were preprocessed based on the radar and image data. A rough detection algorithm calculates and segments the saliency image to obtain regions of interest (ROI). The computational complexity of the network training and prediction was reduced by setting the image pixels outside the ROI. In the detection stage based on deep learning, a lightweight network structure based on ResNet18 was designed to fuse the saliency images at different convolution depths. In the post-decoding stage, the calibration points are fused for local non-maximum suppression (NMS). In contrast to typical detection methods, the proposed method improves the detection speed by removing redundant pixels and local NMS, increasing the detection accuracy by fusing the feature information of the radar and camera in multiple stages. The experimental results indicate that compared with the latest lightweight network (YOLOv8n), the detection accuracy is increased by 10.51%, and the FPS increased by 44.43%. Compared with the latest YOLOv8s and YOLOv9m models, the FPS is increased by 3.0X-4.4X. The field programmable gate array (FPGA) implementation achieves a performance of 60.0 FPS. This is an improvement compared with typical methods, demonstrating that the proposed method is more effective than state-of-the-art models. [ABSTRACT FROM AUTHOR]

Human gaze is a crucial cue used in various applications such as human-robot interaction, autonomous driving, and virtual reality. Recently, convolution neural network (CNN) approaches have made notable progress in predicting gaze angels. However, estimating accurate gaze direction in-the-wild is still a challenging problem due to the difficulty of obtaining the most crucial gaze information that exists in the eye area which constitutes a small part of the face images. In this paper, we introduce a novel two-branch CNN architecture with a multi-loss approach to estimate gaze angles (pitch and yaw) from face images. Our approach utilizes separate fully connected layers for each gaze angle prediction, allowing explicit learning of discriminative features and emphasizing the distinct information associated with each gaze angle. Moreover, we adopt a multi-loss approach, incorporating both classification and regression losses. This allows for joint optimization of the combined loss for each gaze angle, resulting in improved overall gaze performance. To evaluate our model, we conduct experiments on three popular datasets collected under unconstrained settings: MPIIFaceGaze, Gaze360, and RT-GENE. Our proposed model surpasses current state-of-the-art methods and achieves state-of-the-art performance on all three datasets, showcasing its superior capability in gaze estimation. [ABSTRACT FROM AUTHOR]

Environmental mapping serves as a crucial element in Simultaneous Localization and Mapping (SLAM) algorithms, playing a pivotal role in ensuring the accurate representation necessary for autonomous robot navigation guided by SLAM. Current SLAM systems predominantly rely on grid-based map representations, encountering challenges such as measurement discretization for cell fitting and grid map interpolation for online posture prediction. Splines present a promising alternative, capable of mitigating these issues while maintaining computational efficiency. This paper delves into the efficiency disparities between B-Spline surface mapping and discretized cell-based approaches, such as grid mapping, within indoor environments. B-Spline Online SLAM and FastSLAM, utilizing Rao-Blackwellized Particle Filter (RBPF), are employed to achieve range-based mapping of the unknown 2D environment. The system incorporates deep learning networks in the B-Spline curve estimation process to compute parameterizations and knot vectors. The research implementation utilizes the Intel Research Lab benchmark dataset to conduct a comprehensive qualitative and quantitative analysis of both approaches. The B-Spline surface approach demonstrates significantly superior performance, evidenced by low error metrics, including an average squared translational error of 0.0016 and an average squared rotational error of 1.137. Additionally, comparative analysis with Vision Benchmark Suite demonstrates robustness across different environments, highlighting the effectiveness of B-Spline SLAM for real-world applications. [ABSTRACT FROM AUTHOR]