Feature Fusion Research Articles

Improved YOLOv8n-based bridge crack detection algorithm under complex background conditions

Deep learning-based image processing methods are commonly used for bridge crack detection. Aiming at the problem of missed detections and false positives caused by light, stains, and dense cracks during detection, this paper proposes a bridge crack detection algorithm based on the improved YOLOv8n model. Firstly, enhancing the model’s feature extraction capabilities by incorporating the global attention mechanism into the Backbone and Neck to gather additional crack characterization information. And optimizing the original feature fusion model through Gam-Concat to enhance the feature fusion effect. Subsequently, in the FPN-PAN structure, replacing the original upsample module with DySample promotes the full fusion of high- and low-resolution feature information, enhancing the detection capability for cracks of different scales. Finally, adding MPDIoU to the Head to optimize the bounding box function loss, enhancing the model’s ability to evaluate the overlap of dense cracks and better reflecting the spatial relationships between the cracks. In ablation and comparison experiments, the improved model achieved increases of 3.02%, 3.39%, 2.26%, and 0.81% in mAP@0.5, mAP@0.5:0.95, precision, and recall, respectively, compared to the original model. And the detection accuracy is significantly higher than other comparative models. It has practical application value in bridge inspection projects.

LViT-Net: a domain generalization person re-identification model combining local semantics and multi-feature cross fusion

In the task of domain generalization person re-identification (ReID), pedestrian image features exhibit significant intra-class variability and inter-class similarity. Existing methods rely on a single feature extraction architecture and struggle to capture both global context and local spatial information, resulting in weaker generalization to unseen domains. To address this issue, an innovative domain generalization person ReID method–LViT-Net, which combines local semantics and multi-feature cross fusion, is proposed. LViT-Net adopts a dual-branch encoder with a parallel hierarchical structure to extract both local and global discriminative features. In the local branch, the local multi-scale feature fusion module is designed to fuse local feature units at different scales to ensure that the fine-grained local features at various levels are accurately captured, thereby enhancing the robustness of the features. In the global branch, the dual feature cross fusion module fuses local features and global semantic information, focusing on critical semantic information and enabling the mutual refinement and matching of local and global features. This allows the model to achieve a dynamic balance between detailed and holistic information, forming robust feature representations of pedestrians. Extensive experiments demonstrate the effectiveness of LViT-Net. In both single-source and multi-source comparison experiments, the proposed method outperforms existing state-of-the-art methods.

Deep learning-based hippocampus asymmetry assessment for Alzheimer's disease diagnosis.

The symmetry of the brain hippocampus may be disrupted by natural aging and neurodegenerative diseases. Currently, clinical studies on hippocampus asymmetry are limited to subjective visual evaluation and rough volume measurements, lacking quantitative standards. This paper proposes a quantitative assessment method of the hippocampus asymmetry based on deep learning, named DeepHAA (Deep Learning-based Hippocampus Asymmetry Assessment). The DeepHAA model extracts feature representations of left and right hippocampus structures in MRI images and achieved feature fusion through a cross-attention mechanism. A quantitative assessment method is proposed based on the distance between the multimodal embedding of the input sample and the reference embeddingspace. The experimental dataset of this paper included MRI scans of 199 subjects, including 53 normal cognition (NC), 71 mild cognitive impairment (MCI) and 33 Alzheimer's disease (AD). The experimental results show that DeepHAA model can effectively identify and distinguish the NC, MCI, andAD. The proposed deep learning method integrates asymmetric information about hippocampus structure into the diagnosis of AD and has potential clinical applicationvalue.

Remaining useful life prediction of rolling bearings based on multi-wavelet-time convolution transformer

Abstract Aiming at the problems of insufficient local information extraction ability of traditional Transformer and the difficulty of capturing temporal features from the degradation process in the remaining useful life (RUL) prediction, a novel rolling bearing RUL prediction method based on multi-wavelet time convolution transformer (MWTCT) is proposed. First, a temporal attention unit (TAU) is designed to capture the relevant degradation state features at different moments, and a multiple wavelet convolution layer (MWCL) is also designed to acquire the time-frequency features from vibration signal. Then, in order to take into account the sensitivity between different degradation features in the process of feature fusion, cross-attention mechanism are employed to generate salient feature. Finally, to comprehensively capture the global and local degradation features from the extracted features, a multiscale convolutional transformer (MCT) is constructed by combining the multiscale TCN (MSTCN) and the multi-head attention mechanism. Two sets of experiments are conducted on the PHM 2012 datasets to verify the rationality and effectiveness of the proposed algorithm, the results indicate that performance of our method is superior to the existing SOTA methods.

MIT-SAM: Medical Image-Text SAM with Mutually Enhanced Heterogeneous Features Fusion for Medical Image Segmentation.

In recent times, leveraging lesion text as supplementary data to enhance the performance of medical image segmentation models has garnered attention. Previous approaches only used attention mechanisms to integrate image and text features, while not effectively utilizing the highly condensed textual semantic information in improving the fused features, resulting in inaccurate lesion segmentation. This paper introduces a novel approach, the Medical Image-Text Segment Anything Model (MIT-SAM), for text-assisted medical image segmentation. Specifically, we introduce the SAM-enhanced image encoder and a Bert-based text encoder to extract heterogeneous features. To better leverage the highly condensed textual semantic information for heterogeneous feature fusion, such as crucial details like position and quantity, we propose the image-text interactive fusion (ITIF) block and self-supervised text reconstruction (SSTR) method. The ITIF block facilitates the mutual enhancement of homogeneous information among heterogeneous features and the SSTR method empowers the model to capture crucial details concerning lesion text, including location, quantity, and other key aspects. Experimental results demonstrate that our proposed model achieves state-of-the-art performance on the QaTa-COV19 and MosMedData+ datasets. The code of MIT-SAM is available at https://github.com/jojodan514/MIT-SAM.

A target detection model HR-YOLO for advanced driver assistance systems in foggy conditions

To improve the accuracy and real-time performance of detection algorithms in Advanced Driver Assistance Systems (ADAS) under foggy conditions, this paper introduces HR-YOLO, an improved YOLO-based model specifically designed for vehicle and pedestrian detection. To enhance detection performance under complex meteorological conditions, several critical modules have been optimized. First, the Efficient High-Precision Defogging Network (EHPD-Net) is introduced to strengthen feature extraction. Inspired by the Global Attention Mechanism (GAM), the Enhanced Global-Spatial Attention (EGSA) module is incorporated to effectively improve the detection of small targets in foggy conditions. Second, the Depth-Normalized Defogging Network (DND-Net) is applied to enhance image quality. Additionally, the Dynamic Sample (DySample) module is integrated into the neck network, complemented by optimizations to the convolution and C2f modules, which significantly improve feature fusion efficiency. Furthermore, The Wise Intersection over Union (WIoU) loss function is introduced to enhance target localization accuracy. The robustness and accuracy of HR-YOLO were validated through experiments on two foggy weather datasets: RTTS and Foggy Cityscapes. The results indicate that HR-YOLO achieved a mean Average Precision (mAP) of 79.8% on the RTTS dataset, surpassing the baseline by 5.9%. On the Foggy Cityscapes dataset, it achieved an mAP of 49.5%, representing a 9.7% improvement over the baseline. This model serves as an effective solution for target detection tasks under foggy conditions and establishes a foundation for future advancements in this field.

Prediction of Recurrence using a Stacked Denoising Autoencoder and Multifaceted Feature Analysis of Pretreatment MRI in Patients with Nasopharyngeal Carcinoma.

Strategies for predicting the recurrence of nasopharyngeal carcinoma need to be further development and validation. We developed a recurrence prediction model based on the fusion of multi-omics features from pre-treatment conventional magnetic resonance sequences (CE-T1W) in nasopharyngeal carcinoma (NPC) patients to predict posttreatment recurrence. We employed a deep unsupervised stacked denoising autoencoder (stacked denoising autoencoder, SDAE) and multi-omics feature fusion method to develop an NPC recurrence prediction model. Data and magnetic resonance images from 184 patients with newly diagnosed nasopharyngeal carcinoma (NPC), confirmed by pathological examination and who underwent radical comprehensive treatment, were collected. Propensity score matching (relapse: no recurrence = 1:1) was used to balance clinical factors that might influence recurrence, resulting in 136 matched cases. SDAE was utilized to extract deep features, combined with the fusion features of radiomics (Radiomics) features and clinical features, using support vector machine (SVM), multilayer perceptron (MLP), logistic regression (LR) and random forest (RF) machine learning methods to build models. The mean area under the curve (AUC), accuracy, sensitivity, and specificity of each model were compared to evaluate their performance in predicting recurrence. After parameter adjustment, 12 machine learning models based on different fusion features were developed. Model 1 (Radiomics+AutoEncoder+Clinical+SVM) achieved better prediction performance, with a mean AUC, accuracy, sensitivity, and specificity of 0.89 (95% CI: 0.84- 0.93), 81.5%, 67.3% and 97.9%, respectively. Model 2 (Radiomics+ AutoEncoder+SVM), Model 3 (Radiomics+SVM), Model 4 (Radiomics+ AutoEncoder+Clinical+MLP), Model 5 (Radiomics+Auto Encoder+MLP), Model 6 (Radiomics+MLP), Model 7 (Radiomics+AutoEncoder+Clinical+LR), Model 8 (Radiomics+AutoEncoder+LR), Model 9 (Radiomics+LR), Model 10 (Radiomics+Auto Encoder+Clinical+RF), Model 11 (Radiomics+AutoEncoder+RF), Model 12 (Radiomics+RF) achieved AUCs of 0.87, 0.87, 0.82, 0.80, 0.82, 0.82, 0.78, 0.80, 0.81, 0.80, and 0.82, respectively. The SVM model for predicting the recurrence of nasopharyngeal carcinoma based on radiomics+autoEncoder+clinical fusion features established by CE-T1WI of nasopharyngeal carcinoma patients before treatment, achieved good predictive performance and is relatively reliable, which can provide more information and help for clinical diagnosis and treatment decisions and interventions.

Multimodal Fall Detection Using Spatial–Temporal Attention and Bi-LSTM-Based Feature Fusion

Human fall detection is a significant healthcare concern, particularly among the elderly, due to its links to muscle weakness, cardiovascular issues, and locomotive syndrome. Accurate fall detection is crucial for timely intervention and injury prevention, which has led many researchers to work on developing effective detection systems. However, existing unimodal systems that rely solely on skeleton or sensor data face challenges such as poor robustness, computational inefficiency, and sensitivity to environmental conditions. While some multimodal approaches have been proposed, they often struggle to capture long-range dependencies effectively. In order to address these challenges, we propose a multimodal fall detection framework that integrates skeleton and sensor data. The system uses a Graph-based Spatial-Temporal Convolutional and Attention Neural Network (GSTCAN) to capture spatial and temporal relationships from skeleton and motion data information in stream-1, while a Bi-LSTM with Channel Attention (CA) processes sensor data in stream-2, extracting both spatial and temporal features. The GSTCAN model uses AlphaPose for skeleton extraction, calculates motion between consecutive frames, and applies a graph convolutional network (GCN) with a CA mechanism to focus on relevant features while suppressing noise. In parallel, the Bi-LSTM with CA processes inertial signals, with Bi-LSTM capturing long-range temporal dependencies and CA refining feature representations. The features from both branches are fused and passed through a fully connected layer for classification, providing a comprehensive understanding of human motion. The proposed system was evaluated on the Fall Up and UR Fall datasets, achieving a classification accuracy of 99.09% and 99.32%, respectively, surpassing existing methods. This robust and efficient system demonstrates strong potential for accurate fall detection and continuous healthcare monitoring.

Multi domain feature fusion model for EEG based biomedical signal analysis in depression detection with a reduced electrode configuration

Multi domain feature fusion model for EEG based biomedical signal analysis in depression detection with a reduced electrode configuration

Intelligent livestock farming: monitoring pig behavior with enhanced YOLOv5 for spatial temporal feature fusion

PurposeThe livestock industry is undergoing a critical transition to intensive, large-scale farming. Intelligent monitoring technologies are essential for improving epidemic early warning systems, reducing breeding costs, and promoting sustainable production. This study aimed to develop a novel pig behavior recognition method using advanced computer vision technology to support intelligent livestock farming.Design/methodology/approachThe YOLOv5 model was utilized to achieve contactless and efficient monitoring of daily pig activities. The study enhanced the YOLOv5 model by improving its input mechanism, backbone network and by incorporating the shuffle attention module. These modifications significantly improved the ability of the model to capture and interpret the spatiotemporal features of pig behavior.FindingsThe experimental results demonstrate that compared with the original YOLOv5 model, the Precision, Recall, mAP@0.5 and mAp@0.5:0.95 of the proposed model has improved by 3.0%, 2.3%, 2.6% and 10.5%, respectively. These findings showcase the model’s effectiveness and potential for real-world applications in intelligent livestock farming, and highlight the feasibility of employing advanced computer vision models to enhance monitoring and management in animal farming environments.Originality/valueThis study presents a novel approach to pig behavior recognition by integrating cutting-edge computer vision techniques with YOLOv5 enhancements. This study contributes to the field by addressing the challenges of spatiotemporal feature extraction and demonstrating the practical application of these methods in intelligent livestock farming. Future research directions include generalization to other animal species, integration with other sensor data, teal-time monitoring and decision support and application in wildlife and laboratory animal research, thus further advancing the intelligent breeding industry.

PBD-YOLO: Dual-Strategy Integration of Multi-Scale Feature Fusion and Weak Texture Enhancement for Lightweight Particleboard Surface Defect Detection

Surface defect detection plays an important role in particleboard quality control. But it still faces challenges in detecting coexisting multi-scale defects and weak texture defects. To address these issues, this study proposed PBD-YOLO (Particleboard Defect-You Only Look Once), a lightweight YOLO-based algorithm with multi-scale feature fusion and weak texture enhancement capabilities. In order to improve the ability of the algorithm to extract weak texture features, the SPDDEConv (Space to Depth and Difference Enhance Convolution) module was introduced in this study, which reduced the loss of information in the down-sampling process through space-to-depth transformation and enhanced the edge information of weak texture defects through difference convolution. This approach improved the mAP (mean average precision) of weakly featured but edge-sensitive defects (such as scratches) by as much as 20.9%. In order to improve the algorithm’s ability to detect multi-scale defects, this study introduced the ShareSepHead (Share Separated Head) and C2f_SAC (C2f module with Switchable Atrous Convolution) modules. ShareSepHead fused feature maps from different scales of the neck network by adding a convolutional layer with shared weights, and the C2f_SAC module adaptively fused multi-rate receptive fields through a switching mechanism. The synergistic effect of ShareSepHead and C2f_SAC improved the detection accuracy of multi-scale defects by 10.6–20.8%. The experimental results demonstrated that PBD-YOLO achieved 85.6% mAP at 50% intersection over union (IoU) and 81.4% recall, surpassing YOLOv10 by 5.5% and 13%, respectively, while reducing parameters by 11.3%. In summary, it could be better to meet the need of accurately detecting surface defects on particleboard.

Assessment of Landslide Susceptibility Based on ReliefF Feature Weight Fusion: A Case Study of Wenxian County, Longnan City

The Longnan mountainous area, characterized by its complex geological structure and fragile geological environment, is one of the four major regions in China prone to geological disasters. Previous studies have employed traditional evaluation methods to assess landslide susceptibility in the Longnan mountainous area. However, these traditional methods are often subjective, and their accuracy and efficiency are difficult to guarantee. This study, supported by GIS technology, focuses on Wen County in Longnan City, a region frequently affected by landslide disasters. Based on 260 collected landslide disaster points, the study combines the ReliefF model to evaluate and zone landslide susceptibility in Wen County, Longnan City, based on feature contribution values. The lithology and rainfall factors have significant impacts on geological disasters, respectively. Areas along rivers and roads, with loose soil, heavy rainfall, steep slopes, and dense vegetation, are more prone to landslide disasters due to the combined effects of natural factors and human activities. This study also uses the receiver operating characteristic (ROC) curve to validate the accuracy of the evaluation results. The area under the curve (AUC) for the ReliefF feature fusion method is 0.853, which is higher than the 0.838 obtained from the information value method. The ReliefF method demonstrates excellent performance in landslide susceptibility evaluation, offering better predictive capability at a lower computational cost, thus achieving a balance between accuracy and efficiency. This approach can provide valuable references for rapid decision-making by relevant geological disaster prevention and management departments.

Lightweight multi-scale convolution with blended feature attention for facial expression recognition in the wild

Abstract Facial expression recognition (FER) has achieved excellent performance in recent years under the controlled scenarios through deep learning methods. However, the accurate recognition of facial expression in the wild conditions with occlusion, pose changes, and uneven lighting still a challenging problem, not to mention the problem of limited computing resources faced by the growing size of proposed network models. To solve these problems, this paper proposes a multi-scale network based on lightweight convolution (MLC-Net), aiming to improve the recognition accuracy of FER in real-world environments while significantly reducing the number of parameters. In MLC-Net, image shallow features are extracted for global and local blocks through pre-extracted blocks. The global feature extraction block uses a mixed washing network as the basis of the Multi-scale Module, reducing the its parameters and computational complexity when extracting different levels of semantic information. Meanwhile, the improved Efficient Lightweight Channel-spatial Attention Module (ELAM) is used to enhance the feature fusion ability of the Multi-scale Module. The local feature extraction block utilizes convolutional groups and Lightweight Spatial Attention Module (LSAM) to extract and enhance local features, guiding the network to pay attention to regions with significant features, and proposes a Local Relationship Transformer (LRT), through which a multi-head attention mechanism is used to establish connections between regions, thus further enhancing the ability to recognize complex expressions. The effectiveness of the proposed MLC-Net is validated on multiple in the wild FER datasets, and the results show that MLC-Net can achieve a good balance between recognition accuracy and network lightweighting, providing a promised solution for practical application of FER.

A simple monocular depth estimation network for balancing complexity and accuracy

Monocular depth estimation plays a crucial role in many downstream visual tasks. Although research on monocular depth estimation is relatively mature, it commonly involves strategies that entail increasing both the computational complexity and the number of parameters to achieve superior performance. Particularly in practical applications, enhancing the accuracy of depth prediction while ensuring computational efficiency remains a challenging issue. To tackle this challenge, we propose a novel and simple depth estimation model called SimMDE, which treats monocular depth estimation as an ordinal regression problem. Beginning with a baseline encoder, our model is equipped with a Deformable Cross-Attention Feature Fusion (DCF) decoder with sparse attention. This decoder efficiently integrates multi-scale feature maps, markedly reducing the quadratic complexity of the Transformer model. For the extraction of finer local features, we propose a Local Multi-dimensional Convolutional Attention (LMC) module. Meanwhile, we propose a Wavelet Attention Transformer (WAT) module to achieve pixel-level precise classification of images. Furthermore, we also conduct extensive experiments on two widely recognized depth estimation benchmark datasets: NYU and KITTI. The experimental findings unequivocally demonstrate that our model attains exceptional accuracy in depth estimation while upholding high computational efficiency. Remarkably, our framework SimMDE, extending from AdaBins, demonstrates enhancements, resulting in substantial improvements of 11.7% and 10.3% in the absolute relative error (AbsRel) on the NYU and KITTI datasets, respectively, with fewer parameters.

Anomaly detection in pneumatic control valves using stacked re-optimized convolutional autoencoder and product quantization under imbalanced data conditions

Abstract Reliable operation of pneumatic control valves is essential for the safety and efficiency of industrial control systems. Existing data-driven anomaly detection methods are typically based on datasets containing faults from faulty valves. However, the scarcity of abnormal samples and tedious or time-consuming of data labeling in practice often lead to imbalanced datasets and unlabeled data, which pose challenges for the application of these methods. To address this problem, we propose a novel anomaly detection method for pneumatic control valves based on stacked re-optimized convolutional autoencoder (S-RCAE) and product quantization (PQ) technology. S-RCAE is designed as a feature extraction network to obtain the latent and essential feature representations of anomalies in the monitoring data. It is stacked by placing a sample selection layer with the DBSCAN clustering algorithm between the two CAEs. Furthermore, the two CAEs are modified by inserting a convolutional block attention module between every two layers. During the training of S-RCAE, the idea of few-shot learning and an improved cost-sensitive function are adopted for the labeling and feature extraction of anomaly samples in unlabeled and class-imbalanced datasets. Then, a feature fusion mechanism is implemented for the resulting representation from S-RCAE. The PQ is further introduced to enable dimensionality reduction and storage of the fused features and the accurate identification and assessment of anomalies using quantitative scores. The experimental results show that the proposed method can accurately identify and quantify the abnormal states of the valve from the unlabeled and class-imbalanced datasets.

High-throughput method for improving rice AGB estimation based on UAV multi-source remote sensing image feature fusion and ensemble learning

IntroductionThe rapid and non-destructive estimation of rice aboveground biomass (AGB) is vital for accurate growth assessment and yield prediction. However, vegetation indices (VIs) often suffer from saturation due to high canopy coverage and vertical organs, limiting their accuracy across multiple growth stages. Therefore, this study utilizes UAV-acquired RGB and multi-spectral (MS) images during several critical rice stages to explore the potential of multi-source data fusion for accurately and cost-effectively estimating rice AGB.MethodsHigh-frequency texture features were extracted from RGB images using discrete wavelet transform (DWT), while low-order color moments in RGB and Lab color spaces were calculated. VIs were derived from MS images. Feature selection combined statistical analysis and modeling techniques, with collinearity removed through the Variance Inflation Factor (VIF). The relationships between AGB and the selected features were then analyzed using multiple fitting functions. Both single-type and multi-type features were used to develop individual and ensemble machine learning (ML) models for rice AGB estimation.ResultsThe findings indicate that: (i) Single-type features result in significant errors and low accuracy within the same sensor, but multi-feature fusion improves performance. (ii) Fusing RGB and MS image features enhances AGB estimation accuracy over single-sensor features. (iii) Ensemble ML models outperform individual models, providing higher accuracy and stability, with the best model achieving an R2 of 0.8564 and RMSE of 169.32 g/m2.DiscussionThis study demonstrates that multi-source UAV image feature fusion with ensemble learning effectively leverages complementary data strengths, offering an efficient solution for monitoring rice AGB across growth stages.

A Novel Few-Shot Cross-Domain Fault Diagnosis Method Enhanced by Feature Fusion and ReLU-KAN

Abstract The increasing complexity of industrial machinery necessitates reliable fault diagnosis systems to prevent safety risks and operational disruptions. Rolling bearings, as critical components, require precise fault detection, yet existing deep learning models struggle with generalization under data scarcity and cross-domain distribution shifts. To address this, a novel few-shot cross-domain fault diagnosis method enhanced by feature fusion and ReLU-KAN (FFRK), is proposed. This method comprises a feature extraction and fusion module and a feature dimensionality reduction module. Within the prototypical network framework, EfficientViT is employed for feature extraction to capture the multi-scale characteristics of bearing fault signals, while Adaptive Average Pooling is utilized for feature fusion to comprehensively extract fault features. The ReLU-KAN module further enhances model robustness through feature dimensionality reduction. Extensive experiments on three bearing datasets demonstrate FFRK’s superior performance in cross-domain scenarios of speed variation and size variation, achieving state-of-the-art accuracy with minimal training samples. The contributions include: (1) A lightweight multi-scale feature extractor combining EfficientViT and Adaptive Average Pooling; (2) A ReLU-KAN module balancing nonlinear representation and computational efficiency; (3) Comprehensive validation in challenging industrial scenarios, confirming FFRK’s potential for real-time fault diagnosis.

Remote SensingImage Segmentation Using Vision Mamba and Multi-Scale Multi-Frequency Feature Fusion

Rapid advancements in remote sensing (RS) imaging technology have heightened the demand for the precise and efficient interpretation of large-scale, high-resolution RS images. Although segmentation algorithms based on convolutional neural networks (CNNs) or Transformers have achieved significant performance improvements, the trade-off between segmentation precision and computational complexity remains a key limitation for practical applications. Therefore, this paper proposes CVMH-UNet—a hybrid semantic segmentation network that integrates the Vision Mamba (VMamba) framework with multi-scale feature fusion—to achieve high-precision and relatively efficient RS image segmentation. CVMH-UNet comprises the following two core modules: the hybrid visual state space block (HVSSBlock) and the multi-frequency multi-scale feature fusion block (MFMSBlock). The HVSSBlock integrates convolutional branches to enhance local feature extraction while employing a cross 2D scanning method (CS2D) to capture global information from multiple directions, enabling the synergistic modeling of global and local features. The MFMSBlock introduces multi-frequency information via 2D Discrete Cosine Transform (2D DCT) and extracts multi-scale local details through point-wise convolution, thereby optimizing refined feature fusion in skip connections between the encoder and decoder. Experimental results on benchmark RS datasets demonstrate that CVMH-UNet achieves state-of-the-art segmentation accuracy with optimal computational efficiency, surpassing existing advanced methods.

Cross-domain gesture activity recognition via dual-path prototypical network using Wi-Fi signals

Abstract With the rapid advancements in human-computer interaction, gesture activity recognition based on Wi-Fi sensing has emerged as a prominent research area in artificial intelligence. However, cross-domain gesture activity recognition faces significant challenges due to the differences in data distribution between domains. Specifically, models trained in the source domain often experience a significant drop in recognition performance when applied to the target domain. To address this issue, this paper proposes a cross-domain recognition method based on a dual-path prototypical network, FAPNet (Fusion of Amplitude and Phase Network). Compared to traditional single-path prototypical networks, which are prone to information loss due to single-modal feature extraction, and fixed-weight amplitude-phase feature fusion strategies, which struggle to adapt to the nonlinear shifts in cross-domain feature distributions, the proposed method constructs a dual-stream heterogeneous learning architecture that significantly enhances cross-domain recognition performance. Specifically, FAPNet extracts amplitude and phase features through independent branches and further enhances key gesture representations by incorporating a multi-scale spatiotemporal enhancement module; building on this, this paper designs a hierarchical prototype optimization mechanism, which generates global prototypes through self-supervised clustering and incorporates an intra-class feature cosine weighting strategy to effectively enhance the domain adaptability of prototype representations; furthermore, to achieve adaptive fusion of amplitude and phase features, this paper constructs a dual-metric mechanism that evaluates modality uncertainty through feature distance correlation analysis, enabling the autonomous adaptation of amplitude-phase weights. Experiments show that the proposed method achieves high classification accuracy in intra-domain tasks and demonstrates strong cross-domain recognition performance when training data is sufficient or cross-domain samples are limited. Compared to the cross-scenario performance of single-path prototypical networks, it improves by 2.66%, providing an effective solution for cross-domain gesture recognition.

DCATNet: polyp segmentation with deformable convolution and contextual-aware attention network

Polyp segmentation is crucial in computer-aided diagnosis but remains challenging due to the complexity of medical images and anatomical variations. Current state-of-the-art methods struggle with accurate polyp segmentation due to the variability in size, shape, and texture. These factors make boundary detection challenging, often resulting in incomplete or inaccurate segmentation. To address these challenges, we propose DCATNet, a novel deep learning architecture specifically designed for polyp segmentation. DCATNet is a U-shaped network that combines ResNetV2-50 as an encoder for capturing local features and a Transformer for modeling long-range dependencies. It integrates three key components: the Geometry Attention Module (GAM), the Contextual Attention Gate (CAG), and the Multi-scale Feature Extraction (MSFE) block. We evaluated DCATNet on five public datasets. On Kvasir-SEG and CVC-ClinicDB, the model achieved mean dice scores of 0.9351 and 0.9444, respectively, outperforming previous state-of-the-art (SOTA) methods. Cross-validation further demonstrated its superior generalization capability. Ablation studies confirmed the effectiveness of each component in DCATNet. Integrating GAM, CAG, and MSFE effectively improves feature representation and fusion, leading to precise and reliable segmentation results. These findings underscore DCATNet’s potential for clinical application and can be used for a wide range of medical image segmentation tasks.