Channel Attention Research Articles

Non-Contact Blood Pressure Estimation from Radar Signals by a Stacked Deformable Convolution Network.

This study introduces a contactless blood pressure monitoring approach that combines conventional radar signal processing with novel deep learning architectures. During the preprocessing phase, datasets suitable for synchronization are created by integrating Kalman filtering, multiscale bandpass filters, and a periodic extraction method in the time domain. These data comprise data on chest micro variations, encapsulating a complex array of physiological and biomedical information reflective of cardiac micromotions. The Radar-based Stacked Deformable convolution Network (RSD-Net) integrates channel and spatial self attention mechanisms within a deformable convolutional framework to enhance feature extraction from radar signals. The network architecture systematically employs deformable convolutions for initial deep feature extraction from individual signals. Subsequently, continuous blood pressure estimation is conducted using self attention mechanisms on feature map from single source coupled with multi-feature map channel attention. The performance of model is corroborated via the open-source dataset procured using a non-invasive 24GHz six-port continuous wave radar system. The dataset, encompassing readings from 30 healthy individuals subjected to diverse conditions including rest, the Valsalva maneuver, apnea, and tilt-table examinations. It serves to substantiate the validity and resilience of the proposed method in the non-contact assessment of continuous blood pressure. Evaluation metrics reveal Pearson correlation coefficients of 0.838 for systolic and 0.797 for diastolic blood pressure predictions. The Mean Error (ME) and Standard Deviation (SD) for systolic and diastolic blood pressure measurements are -0.32 ±6.14mmHg and -0.20 ±5.50mmHg, respectively. The ablation study assesses the contribution of different structural components of the RSD-Net, validating their significance in the overall of model performance.

Self-supervised monocular visual odometry based on cross-correlation

Visual odometry constitutes a critical component in enabling autonomous navigation. However, the existing methods are limited by the feature extraction and matching accuracy, and cannot show good real-time performance while combining accuracy and robustness. In this paper, we propose a novel monocular visual odometry framework based on cross-correlation. The framework starts with a parameter-sharing Siamese network to build feature extractors that can simultaneously process multiple images as inputs. Moreover, we design cross-correlation modules and define a cross-correlation matrix to describe the strength of correlation between different parts of the input feature maps, reflecting the rotational and translational transformations of the input images. Furthermore, a novel loss function is introduced to impose constraints on the network. Additionally, a fully convolutional network is designed for pose estimation, computing poses alterations from the structure of the cross-correlation matrix. Channel attention and spatial attention mechanisms are introduced to improve the performance. More importantly, our method innovatively uses time intervals as labels, enables self-supervised training, and relies only on a monocular camera. Experimental results on the KITTI visual odometry dataset and the Oxford Robotcar Dataset show that our method produces competitive performance, demonstrating the superiority of the proposed method.

Fault Diagnosis Method for Railway Signal Equipment Based on Data Enhancement and an Improved Attention Mechanism

Railway signals’ fault text data contain a substantial amount of expert maintenance experience. Extracting valuable information from these fault text data can enhance the efficiency of fault diagnosis for signal equipment, thereby contributing to the advancement of intelligent railway operations and maintenance technology. Considering that the characteristics of different signal equipment in actual operation can easily lead to a lack of fault data, a fault diagnosis method for railway signal equipment based on data augmentation and an improved attention mechanism (DEIAM) is proposed in this paper. Firstly, the original fault dataset is preprocessed based on data augmentation technology and retained noun and verb operations. Then, the neural network is constructed by integrating a bidirectional long short-term memory (BiLSTM) model with an attention mechanism and a convolutional neural network (CNN) model enhanced with a channel attention mechanism. The DEIAM method can more effectively capture the important text features and sequence features in fault text data, thereby facilitating the diagnosis and classification of such data. Consequently, it enhances onsite fault maintenance experience by providing more precise insights. An empirical study was conducted on a 10-year fault dataset of signal equipment produced by a railway bureau. The experimental results demonstrate that in comparison with the benchmark model, the DEIAM model exhibits enhanced performance in terms of accuracy, precision, recall, and F1.

A maize seed variety identification method based on improving deep residual convolutional network

Seed quality and safety are related to national food security, and seed variety purity is an essential indicator in seed quality detection. This study established a maize seed dataset comprising 5877 images of six different types and proposed a maize seed recognition model based on an improved ResNet50 framework. Firstly, we introduced the ResStage structure in the early stage of the original model, which facilitated the network’s learning process and enabled more efficient information propagation across the network layers. Meanwhile, in the later residual blocks of the model, we introduced both the efficient channel attention (ECA) mechanism and depthwise separable (DS) convolution, which reduced the model’s parameter cost and enabled the capturing of more precise and detailed features. Finally, a Swish-PReLU mixed activation function was introduced globally to improve the overall predictive power of the model. The results showed that our model achieved an impressive accuracy of 91.23% in corn seed classification, surpassing other related models. Compared with the original model, our model improved the accuracy by 7.07%, reduced the loss value by 0.19, and decreased the number of parameters by 40%. The research suggested that this method can efficiently classify corn seeds, holding significant value in seed variety identification.

G2D-Net: Efficient Dehazing with Second-Order Gated Units

Image dehazing aims to reconstruct potentially clear images from corresponding images corrupted by haze. With the rapid development of deep learning-related technologies, dehazing methods based on deep convolutional neural networks have gradually become mainstream. We note that existing dehazing methods often accompany an increase in computational overhead while improving the performance of dehazing. We propose a novel lightweight dehazing neural network to balance performance and efficiency: the g2D-Net. The g2D-Net borrows the design ideas of input-adaptive and long-range information interaction from Vision Transformers and introduces two kinds of convolutional blocks, i.e., the g2D Block and the FFT-g2D Block. Specifically, the g2D Block is a residual block with second-order gated units, which inherit the input-adaptive property of a gated unit and can realize the second-order interaction of spatial information. The FFT-g2D Block is a variant of the g2D Block, which efficiently extracts the global features of the feature maps through fast Fourier convolution and fuses them with local features. In addition, we employ the SK Fusion layer to improve the cascade fusion layer in a traditional U-Net, thus introducing the channel attention mechanism and dynamically fusing information from different paths. We conducted comparative experiments on five benchmark datasets, and the results demonstrate that the g2D-Net achieves impressive dehazing performance with relatively low complexity.

Internet of Things intrusion detection: Research and practice of NSENet and LSTM fusion models

To address the problems of complex environment, limited device computational resources and limited memory resources in the existing IoT, SELSTM, an intrusion detection system composed of NSENet and LSTM fusion based on SENet, is investigated. The NSENet part of the SELSTM system is based on the squeeze-and-excitation network (SENet). The lightweight computational modules NonLocal, SKConv and inverted residuals are fused into SE blocks, and self-attention of Nonlocal is used to improve the local receptive field of feature extraction. The channel attention and spatial attention of each part of the data are strengthened by the use of SKConv To enhance the adaptive convolution ability of the model and ensure the completeness of the information, the properties of the inverted residual structure are used to ensure that the gradient of the model decreases steadily without gradient explosion or disappearance. For the problem of data imbalance, the dataset is randomly resampled using the weight resampling technique to improve the balance of the dataset to ensure that the final detection effect of the model is more effective and generalized, while the data flow is divided into two parts for processing, and the model parameters are optimized using the model gradient optimizer consisting of the optimizer Lion and the optimization function Lookahead. The model extracts the spatial and temporal features of the data through multidimensional extraction to ensure the completeness of the data feature information in multiple dimensions, thus obtaining better detection results. The results of the experiments comparing the SELSTM model with other models on the intrusion dataset show that the intrusion detection model has a higher detection precision and accuracy than the traditional deep learning intrusion detection model, which indicates that the SELSTM has better detection performance properties and better practicality and effectiveness on IoT devices.

A Multiscale Attention Segment Network-Based Semantic Segmentation Model for Landslide Remote Sensing Images

Landslide disasters have garnered significant attention due to their extensive devastating impact, leading to a growing emphasis on the prompt and precise identification and detection of landslides as a prominent area of research. Previous research has primarily relied on human–computer interactions and visual interpretation from remote sensing to identify landslides. However, these methods are time-consuming, labor-intensive, subjective, and have a low level of accuracy in extracting data. An essential task in deep learning, semantic segmentation, has been crucial to automated remote sensing image recognition tasks because of its end-to-end pixel-level classification capability. In this study, to mitigate the disadvantages of existing landslide detection methods, we propose a multiscale attention segment network (MsASNet) that acquires different scales of remote sensing image features, designs an encoder–decoder structure to strengthen the landslide boundary, and combines the channel attention mechanism to strengthen the feature extraction capability. The MsASNet model exhibited an average accuracy of 95.13% on the test set from Bijie’s landslide dataset, a mean accuracy of 91.45% on the test set from Chongqing’s landslide dataset, and a mean accuracy of 90.17% on the test set from Tianshui‘s landslide dataset, signifying its ability to extract landslide information efficiently and accurately in real time. Our proposed model may be used in efforts toward the prevention and control of geological disasters.

Segmentation of the left atrial appendage based on fusion attention.

In clinical practice, the morphology of the left atrial appendage (LAA) plays an important role in the selection of LAA closure devices for LAA closure procedures. The morphology determination is influenced by the segmentation results. The LAA occupies only a small part of the entire 3D medical image, and the segmentation results are more likely to be biased towards the background region, making the segmentation of the LAA challenging. In this paper, we propose a lightweight attention mechanism called fusion attention, which imitates human visual behavior. We process the 3D image of the LAA using a method that involves overview observation followed by detailed observation. In the overview observation stage, the image features are pooled along the three dimensions of length, width, and height. The obtained features from the three dimensions are then separately input into the spatial attention and channel attention modules to learn the regions of interest. In the detailed observation stage, the attention results from the previous stage are fused using element-wise multiplication and combined with the original feature map to enhance feature learning. The fusion attention mechanism was evaluated on a left atrial appendage dataset provided by Liaoning Provincial People's Hospital, resulting in an average Dice coefficient of 0.8855. The results indicate that the fusion attention mechanism achieves better segmentation results on 3D images compared to existing lightweight attention mechanisms.

Semantic Coarse-to-Fine Granularity Learning for Two-Stage Few-Shot Anomaly Detection

Anomaly detection is critical in industrial inspection, where identifying defects significantly impacts product quality and safety. Existing models, primarily based on convolutional neural networks (CNNs), struggle with noise sensitivity and insufficient resolution for fine-grained feature discrimination. To address these issues, we propose a two-stage few-shot anomaly detection network that enhances semantic feature granularity and generalization. The network includes a coarse-grained anomaly detection module, a multi-scale channel attention module, and a fine-grained detection module. The coarse-grained module identifies abnormal regions, serving as the initial filter. The multi-scale channel attention module focuses on anomalous features, enhancing sensitivity to fine-grained characteristics. This step overcomes limitations in discerning subtle yet critical anomalies. The fine-grained detection module refines feature maps, enhancing generalization. Experimental results on the MVTec dataset show an image-level Area under the region of convergence (AUROC) of 92.3% and a pixel-level AUROC of 95.3%, a 1% to 2% improvement over leading FSAD methods.

Multi-hop double-branch network for small sample hyperspectral image classification

ABSTRACT Neural networks have been extensively utilized in the classification of hyperspectral images. The study introduces a novel dual-branch structured network named MGACN for hyperspectral image classification under the condition of small sample sizes. This model a multi-hop graph network (MHGCN) and a graph attention convolutional network (GACNN). Graph Convolutional Network (GCN) integrates multi-hop mechanism to obtain and aggregate information between long-distance non-neighbor nodes. The Graph Attention Network (GAT) combines 2D-CNN by incorporating a noise reduction mechanism, as well as position and channel attention mechanisms. The classification accuracy reaches 98.43%, 99.02%, and 98.55% on the Indian Pines, Salinas, and Pavia University datasets, which is significantly better than that of the comparison model.

EEG motor imagery decoding: A framework for comparative analysis with channel attention mechanisms.

Objective
The objective of this study is to investigate the application of various channel attention mechanisms within the domain of brain-computer interface (BCI) for motor imagery decoding. Channel attention mechanisms can be seen as a powerful evolution of spatial filters traditionally used for motor imagery decoding. This study systematically compares such mechanisms by integrating them into a lightweight architecture framework to evaluate their impact. 
Approach
We carefully construct a straightforward and lightweight baseline architecture designed to seamlessly integrate different channel attention mechanisms. This approach is contrary to previous works which only investigate one attention mechanism and usually build a very complex, sometimes nested architecture. Our framework allows us to evaluate and compare the impact of different attention mechanisms under the same circumstances. The easy integration of different channel attention mechanisms as well as the low computational complexity enables us to conduct a wide range of experiments on four datasets to thoroughly assess the effectiveness of the baseline model and the attention mechanisms. 
Results
Our experiments demonstrate the strength and generalizability of our architecture framework as well as how channel attention mechanisms can improve the performance while maintaining the small memory footprint and low computational complexity of our baseline architecture.
Significance
Our architecture emphasizes simplicity, offering easy integration of channel attention mechanisms, while maintaining a high degree of generalizability across datasets, making it a versatile and efficient solution for EEG motor imagery decoding within brain-computer interfaces.

Multi-Target Feeding-Behavior Recognition Method for Cows Based on Improved RefineMask

Within the current process of large-scale dairy-cattle breeding, to address the problems of low recognition-accuracy and significant recognition-error associated with existing visual methods, we propose a method for recognizing the feeding behavior of dairy cows, one based on an improved RefineMask instance-segmentation model, and using high-quality detection and segmentation results to realize the recognition of the feeding behavior of dairy cows. Firstly, the input features are better extracted by incorporating the convolutional block attention module into the residual module of the feature extraction network. Secondly, an efficient channel attention module is incorporated into the neck design to achieve efficient integration of feature extraction while avoiding the surge of parameter volume computation. Subsequently, the GIoU loss function is used to increase the area of the prediction frame to optimize the convergence speed of the loss function, thus improving the regression accuracy. Finally, the logic of using mask information to recognize foraging behavior was designed, and the accurate recognition of foraging behavior was achieved according to the segmentation results of the model. We constructed, trained, and tested a cow dataset consisting of 1000 images from 50 different individual cows at peak feeding times. The method’s effectiveness, robustness, and accuracy were verified by comparing it with example segmentation algorithms such as MSRCNN, Point_Rend, Cascade_Mask, and ConvNet_V2. The experimental results show that the accuracy of the improved RefineMask algorithm in recognizing the bounding box and accurately determining the segmentation mask is 98.3%, which is higher than that of the benchmark model by 0.7 percentage points; for this, the model parameter count size was 49.96 M, which meets the practical needs of local deployment. In addition, the technologies under study performed well in a variety of scenarios and adapted to various light environments; this research can provide technical support for the analysis of the relationship between cow feeding behavior and feed intake during peak feeding periods.

Attention-Oriented CNN Method for Type 2 Diabetes Prediction

Diabetes is caused by insulin deficiency or impaired biological action, and long-term hyperglycemia leads to a variety of tissue damage and dysfunction. Therefore, the early prediction of diabetes and timely intervention and treatment are crucial. This paper proposes a robust framework for the prediction and diagnosis of type 2 diabetes (T2DM) to aid in diabetes applications in clinical diagnosis. The data-preprocessing stage includes steps such as outlier removal, missing value filling, data standardization, and assigning class weights to ensure the quality and consistency of the data, thereby improving the performance and stability of the model. This experiment used the National Health and Nutrition Examination Survey (NHANES) dataset and the publicly available PIMA Indian dataset (PID). For T2DM classification, we designed a convolutional neural network (CNN) and proposed a novel attention-oriented convolutional neural network (SECNN) through the channel attention mechanism. To optimize the hyperparameters of the model, we used grid search and K-fold cross-validation methods. In addition, we also comparatively analyzed various machine learning (ML) models such as support vector machine (SVM), logistic regression (LR), decision tree (DT), random forest (RF), and artificial neural network (ANN). Finally, we evaluated the performance of the model using performance evaluation metrics such as precision, recall, F1-Score, accuracy, and AUC. Experimental results show that the SECNN model has an accuracy of 94.12% on the NHANES dataset and an accuracy of 89.47% on the PIMA Indian dataset. SECNN models and CNN models show significant improvements in diabetes prediction performance compared to traditional ML models. The comparative analysis of the SECNN model and the CNN model has significantly improved performance, further verifying the advantages of introducing the channel attention mechanism. The robust diabetes prediction framework proposed in this article establishes an effective foundation for diabetes diagnosis and prediction, and has a positive impact on the development of health management and medical industries.

Exploration of Speech and Music Information for Movie Genre Classification

Movie genre prediction from trailers is mostly attempted in a multi-modal manner. However, the characteristics of movie trailer audio indicate that this modality alone might be highly effective in genre prediction. Movie trailer audio predominantly consists of speech and music signals in isolation or overlapping conditions. This work hypothesizes that the genre labels of movie trailers might relate to the composition of their audio component. In this regard, speech-music confidence sequences for the trailer audio are used as a feature. In addition, two other features previously proposed for discriminating speech-music are also adopted in the current task. This work proposes a time and channel Attention Convolutional Neural Network (ACNN) classifier for the genre classification task. The convolutional layers in ACNN learn the spatial relationships in the input features. The time and channel attention layers learn to focus on crucial time steps and CNN kernel outputs, respectively. The Moviescope dataset is used to perform the experiments, and two audio-based baseline methods are employed to benchmark this work. The proposed feature set with the ACNN classifier improves the genre classification performance over the baselines. Moreover, decent generalization performance is obtained for genre prediction of movies with different cultural influences (EmoGDB).

SwinYOLOv7: Robust ship detection in complex synthetic aperture radar images

Using satellite-based SAR (Synthetic Aperture Radar) imagery to detect and track ships is a formidable challenge. However, accurate analysis is hampered by inherent difficulties such as obscured edges, multiple targets, varying dimensions and complex backgrounds. These factors contribute to a poor signal-to-noise ratio. Various artificial intelligence models have been developed to improve these problems, especially with YOLO-based models. To overcome the above challenges and achieve robust performance, this study presents a model called SwinYOLOv7, a novel fusion of the YOLOv7 framework with the Features Pyramid Network and the Swin Transformer using ground-breaking anchor-free detection algorithms. This innovative method aims to increase the accuracy of vessel detection while reducing the impact of background clutter. The proposed model improved by YOLOv7 examined three different datasets in detail: SRSDD-v1.0, HRSID, and SSDD to optimize the performance of the model. The training process was consistently verified using superior recall, precision, and F1-score values, which can be easily compared with previous studies. The results show that the model using the Swin Transformer attention mechanism and using an image size of 640×640 achieves the highest accuracy of 96.59%. Alternative attention mechanisms, including the Squeeze-and-Excitation Network (SEnet), the Convolutional Block Attention Module (CBAM), Channel Attention (CA) and Efficient Channel Attention (ECA), deliver poorer accuracy rates. The combination of YOLOv7 and Swin Transformer yielded encouraging results that enabled the proposed model to outperform the current benchmarking models. Therefore, the proposed model provides a compelling solution to ensure accurate vessel identification in complex search and rescue scenarios.

MUNet++: Multilevel wavelet nested UNet++ demoiréing residual network

When mobile phones and digital cameras are used to capture information on a screen or store scenes with rich textures, an unpleasant moiré phenomenon occurs, which seriously degrades the quality of the image and affects subsequent image processing tasks. Therefore, to remove the moiré patterns, we propose a multilevel wavelet nested UNet++ demoiréing residual network called MUNet++. Unlike these methods based on the RGB domain, our method is performed in the frequency domain. The multilevel discrete wavelet transform (DWT) and the inverse discrete wavelet transform (IDWT) are elegantly embedded into the UNet++ encoder-decoder structure and transformed to the multilevel frequency domain, which captures more moiré features and preserves details of the image. In addition, we design a dual branch efficient pixel fusion attention module (DBEPFAM) that combines multiscale information, channel attention and the proposed pixel attention fusion module (PAFM) to improve the representation of features. In the network layer, a simple nonlinear unit (NU) is designed and embedded in the middle of the network layer to improve the representation capability of the network. Experimental results on two public datasets show that MUNet++ can effectively remove the moiré patterns and outperform existing state-of-the-art techniques.

EEG classification model for virtual reality motion sickness based on multi-scale CNN feature correlation

BackgroundVirtual reality motion sickness (VRMS) is a key issue hindering the development of virtual reality technology, and accurate detection of its occurrence is the first prerequisite for solving the issue. ObjectiveIn this paper, a convolutional neural network (CNN) EEG detection model based on multi-scale feature correlation is proposed for detecting VRMS. MethodsThe model uses multi-scale 1D convolutional layers to extract multi-scale temporal features from the multi-lead EEG data, and then calculates the feature correlations of the extracted multi-scale features among all the leads to form the feature adjacent matrixes, which converts the time-domain features to correlation-based brain network features, thus strengthen the feature representation. Finally, the correlation features of each layer are fused. The fused features are then fed into the channel attention module to filter the channels and classify them using a fully connected network. Finally, we recruit subjects to experience 6 different modes of virtual roller coaster scenes, and collect resting EEG data before and after the task to verify the model. Results: The results show that the accuracy, precision, recall and F1-score of this model for the recognition of VRMS are 98.66 %, 98.65 %, 98.68 %, and 98.66 %, respectively. The proposed model outperforms the current classic and advanced EEG recognition models. SignificanceIt shows that this model can be used for the recognition of VRMS based on the resting state EEG.

SRMA: a dual-branch parallel multi-scale attention network for remote sensing images sea-land segmentation

ABSTRACT The use of deep learning-based high resolution remote sensing image sea and land segmentation method has become prevalent in various fields such as environmental monitoring, resource assessment, and urban planning. However, the segmentation process faces challenges due to the complex coastline and the similarity in spectral colour difference between the sea and land in certain areas. The remote sensing images obtained from satellites often exhibit unclear boundaries between the sea and land, as well as intricate coastlines. Consequently, the segmentation of sea and land becomes a difficult task. The existing sea-land segmentation methods have limitations in accurately segmenting long and narrow waters, as well as areas with low differentiation between sea and land. In this paper, we propose a new sea-land segmentation method called SRMA (Swin-Res-Multi-Attention). SRMA performs parallel feature extraction using a dual-branch parallel: Swin-Transformer branch and ResNet branch. The combination of Swin-Transformer’s self-attention and ResNet’s residual design helps reduce the loss of feature information, thereby avoiding incorrect segmentation in regions with low distinctness. To improve overall stability, we introduce the MCA (Multiscale channel attention module) module, which combines feature information from different scales. The multi-scale pyramid design of MCA module ensures stability and improves segmentation accuracy in narrow waters. Experimental results on a public dataset demonstrate that our model outperforms existing methods. Additionally, we evaluate the model’s performance in complex scenarios by creating a new dataset that includes a complex urban water body and cloud interference. The results showcase the robustness and superiority of our model.

Electroencephalogram-Based ConvMixer Architecture for Recognizing Attention Deficit Hyperactivity Disorder in Children

Attention deficit hyperactivity disorder (ADHD) is a neuro-developmental disorder that affects approximately 5–10% of school-aged children worldwide. Early diagnosis and intervention are essential to improve the quality of life of patients and their families. In this study, we propose ConvMixer-ECA, a novel deep learning architecture that combines ConvMixer with efficient channel attention (ECA) blocks for the accurate diagnosis of ADHD using electroencephalogram (EEG) signals. The model was trained and evaluated using EEG recordings from 60 healthy children and 61 children with ADHD. A series of experiments were conducted to evaluate the performance of the ConvMixer-ECA. The results showed that the ConvMixer-ECA performed well in ADHD recognition with 94.52% accuracy. The incorporation of attentional mechanisms, in particular ECA, improved the performance of ConvMixer; it outperformed other attention-based variants. In addition, ConvMixer-ECA outperformed state-of-the-art deep learning models including EEGNet, CNN, RNN, LSTM, and GRU. t-SNE visualization of the output of this model layer validated the effectiveness of ConvMixer-ECA in capturing the underlying patterns and features that separate ADHD from typically developing individuals through hierarchical feature learning. These outcomes demonstrate the potential of ConvMixer-ECA as a valuable tool to assist clinicians in the early diagnosis and intervention of ADHD in children.

HST-MRF: Heterogeneous Swin Transformer With Multi-Receptive Field for Medical Image Segmentation.

The Transformer has been successfully used in medical image segmentation due to its excellent long-range modeling capabilities. However, patch segmentation is necessary when building a Transformer class model. This process ignores the tissue structure features within patch, resulting in the loss of shallow representation information. In this study, we propose a Heterogeneous Swin Transformer with Multi-Receptive Field (HST-MRF) model that fuses patch information from different receptive fields to solve the problem of loss of feature information caused by patch segmentation. The heterogeneous Swin Transformer (HST) is the core module, which achieves the interaction of multi-receptive field patch information through heterogeneous attention and passes it to the next stage for progressive learning, thus complementing the patch structure information. We also designed a two-stage fusion module, multimodal bilinear pooling (MBP), to assist HST in further fusing multi-receptive field information and combining low-level and high-level semantic information for accurate localization of lesion regions. In addition, we developed adaptive patch embedding (APE) and soft channel attention (SCA) modules to retain more valuable information when acquiring patch embedding and filtering channel features, respectively, thereby improving model segmentation quality. We evaluated HST-MRF on multiple datasets for polyp, skin lesion and breast ultrasound segmentation tasks. Experimental results show that our proposed method outperforms state-of-the-art models and can achieve superior performance. Furthermore, we verified the effectiveness of each module and the benefits of multi-receptive field segmentation in reducing the loss of structural information through ablation experiments and qualitative analysis.