Spatial Attention Module Research Articles

Mask-guided network for finger vein feature extraction and biometric identification

The problems of complex background, low quality of finger vein images, and poor discriminative features have been the bottleneck of feature extraction and finger vein recognition. To this end, we propose a feature extraction algorithm based on the open-set testing protocol. In order to eliminate the interference of irrelevant areas, this paper proposes the idea of segmentation-assisted classification, that is, using the rough mask of the finger vein to constrain the feature learning process so that the network can focus on the vein area and learn greater weight for the vein. Specifically, the feature maps of the shallow layers of the network are first sent to the feature pyramid module to fuse the primary features of different scales, which are then sent to the spatial attention module to obtain the spatial weight map of the image. Based on the results of several classical vein skeleton extraction algorithms, a weighting method is used to obtain a more accurate mask to constrain the learning of the spatial weight map. Finally, a hybrid loss function combining triplet loss and cross-entropy loss is used to reduce the distance between feature vectors of the same categories and increase the distance between feature vectors of different categories in the Euclidean space, thereby improving feature discriminability. Good recognition results were achieved on the three public data sets of SDUMLA, MMCBNU, and FVUSM, and the values of equal error rate (EER) on them are as low as 2.50%, 0.20%, and 0.14%, respectively.

Semantic Segmentation of Corn Leaf Blotch Disease Images Based on U-Net Integrated with RFB Structure and Dual Attention Mechanism

Northern corn leaf blight (NCLB) is caused by a fungus and can be susceptible to the disease throughout the growing period of corn, posing a significant impact on corn yield. Aiming at the problems of under-segmentation, over-segmentation, and low segmentation accuracy in the traditional segmentation model of northern corn leaf blight, this study proposes a segmentation method based on an improved U-Net network model. By introducing a convolutional layer and maximum pooling layer to a VGG19 network, the channel attention module and spatial attention module (CBAM) are fused, and the squeeze excitation (SE) attention mechanism is combined. This enhances image feature decoding, integrates feature maps of each layer, strengthens the feature extraction process, expands the sensory fields and aggregates context information, and reduces the loss of location and dense semantic information caused by the pooling operation. Findings from the study show that the proposed NCLB-Net has significantly improved the MIoU and PA indexes, reaching 92.43% and 94.71%, respectively. Compared with the traditional methods, U-Net, SETR, DAnet, OCnet, PSPNet, etc., the MIoU is improved by 20.81%, 16.10%, 9.79%, 5.27%, and 11.06%, and the PA is improved by 11.49%, 8.18%, 9.54%, 13.11%, and 6.26%, respectively.

Artificial intelligence-assisted magnetic resonance imaging technology in the differential diagnosis and prognosis prediction of endometrial cancer

It aimed to analyze the value of deep learning algorithm combined with magnetic resonance imaging (MRI) in the risk diagnosis and prognosis of endometrial cancer (EC). Based on the deep learning convolutional neural network (CNN) architecture residual network with 101 layers (ResNet-101), spatial attention and channel attention modules were introduced to optimize the model. A retrospective collection of MRI image data from 210 EC patients was used for model segmentation and reconstruction, with 140 cases as the test set and 70 cases as the validation set. The performance was compared with traditional ResNet-101 model, ResNet-101 model based on spatial attention mechanism (SA-ResNet-101), and ResNet-101 model based on channel attention mechanism (CA-ResNet-101), using accuracy (AC), precision (PR), recall (RE), and F1 score as evaluation metrics. Among the 70 cases in the validation set, there were 45 cases of low-risk EC and 25 cases of high-risk EC. Using ROC curve analysis, it was found that the area under the curve (AUC) for the diagnosis of high-risk EC of the proposed model in this article (0.918) was visibly larger as against traditional ResNet-101 (0.613), SA-ResNet-101 (0.760), and CA-ResNet-101 models (0.758). The AC, PR, RE, and F1 values of the proposed model for the diagnosis of EC risk were visibly higher (P < 0.05). In the validation set, postoperative recurrence occurred in 13 cases and did not occur in 57 cases. Using ROC curve analysis, it was found that the AUC for postoperative recurrence prediction of the patients by the proposed model (0.926) was visibly larger as against traditional ResNet-101 (0.620), SA-ResNet-101 (0.729), and CA-ResNet-101 models (0.767). The AC, PR, RE, and F1 values of the proposed model for postoperative recurrence prediction were visibly higher (P < 0.05). The proposed model in this article, assisted by MRI, presented superior performance in diagnosing high-risk EC patients, with higher sensitivity (Sen) and specificity (Spe), and also demonstrated excellent predictive AC in postoperative recurrence prediction.

SAM-ResNet50: A Deep Learning Model for the Identification and Classification of Drought Stress in the Seedling Stage of Betula luminifera

Betula luminifera, an indigenous hardwood tree in South China, possesses significant economic and ecological value. In view of the current severe drought situation, it is urgent to enhance this tree’s drought tolerance. However, traditional artificial methods fall short of meeting the demands of breeding efforts due to their inefficiency. To monitor drought situations in a high-throughput and automatic approach, a deep learning model based on phenotype characteristics was proposed to identify and classify drought stress in B. luminifera seedlings. Firstly, visible-light images were obtained from a drought stress experiment conducted on B. luminifera shoots. Considering the images’ characteristics, we proposed an SAM-CNN architecture by incorporating spatial attention modules into classical CNN models. Among the four classical CNNs compared, ResNet50 exhibited superior performance and was, thus, selected for the construction of the SAM-CNN. Subsequently, we analyzed the classification performance of the SAM-ResNet50 model in terms of transfer learning, training from scratch, model robustness, and visualization. The results revealed that SAM-ResNet50 achieved an accuracy of 1.48% higher than that of ResNet50, at 99.6%. Furthermore, there was a remarkable improvement of 18.98% in accuracy, reaching 82.31% for the spatial transform images generated from the test set images by applying movement and rotation for robustness testing. In conclusion, the SAM-ResNet50 model achieved outstanding performance, with 99.6% accuracy and realized high-throughput automatic monitoring based on phenotype, providing a new perspective for drought stress classification and technical support for B. luminifera-related breeding work.

Recognition of palaeo-glacial troughs with cirques on parts of Tibetan Plateau using multi-modal datasets with deep learning models

Recognition of landforms associated with past glaciation is crucial for understanding past ice dynamics and their relationship to climate. With the development of artificial intelligence technology, deep learning techniques have assisted in the automatic extraction of glacial landforms, but these methods still face problems of low precision and weak transferability. This study proposes a new method named geomorphology-attention DeeplabV3+ (GA-DeeplabV3+) model. This method adds spatial attention and channel attention modules based on the DeeplabV3+ network and utilizes a combination of multi-modal geographic data. Verification shows that the model proposed in this paper significantly enhances the recognition precision of glacial troughs with cirques compared to existing models, reaching a Mean Intersection over Union (MIoU) of 86.2% and a mean Pixel Accuracy (mPA) of 90.64% in the area of the Palaeo-Daocheng ice cap. In addition, validation experiments were conducted in the Peiku Gangri region and the Tenasserim mountains, achieving MIoU scores of 70.09% and 73.28% respectively. This accomplishment represents a vital stride towards automating the extraction of palaeo-glacial landforms, which holds great significance for analyzing the scale and evolution of ancient glaciers.

LSSF-Net: Lightweight segmentation with self-awareness, spatial attention, and focal modulation

Accurate segmentation of skin lesions within dermoscopic images plays a crucial role in the timely identification of skin cancer for computer-aided diagnosis on mobile platforms. However, varying shapes of the lesions, lack of defined edges, and the presence of obstructions such as hair strands and marker colors make this challenge more complex. Additionally, skin lesions often exhibit subtle variations in texture and color that are difficult to differentiate from surrounding healthy skin, necessitating models that can capture both fine-grained details and broader contextual information. Currently, melanoma segmentation models are commonly based on fully connected networks and U-Nets. However, these models often struggle with capturing the complex and varied characteristics of skin lesions, such as the presence of indistinct boundaries and diverse lesion appearances, which can lead to suboptimal segmentation performance. To address these challenges, we propose a novel lightweight network specifically designed for skin lesion segmentation utilizing mobile devices, featuring a minimal number of learnable parameters (only 0.8 million). This network comprises an encoder–decoder architecture that incorporates conformer-based focal modulation attention, self-aware local and global spatial attention, and split channel-shuffle. The efficacy of our model has been evaluated on four well-established benchmark datasets for skin lesion segmentation: ISIC 2016, ISIC 2017, ISIC 2018, and PH2. Empirical findings substantiate its state-of-the-art performance, notably reflected in a high Jaccard index.

TSRDet: A Table Structure Recognition Method Based on Row-Column Detection

As one of the most commonly used and important data carriers, tables have the advantages of high structuring, strong readability and strong flexibility. However, in reality, tables usually present various forms, such as Excel, images, etc. Among them, the information in the table image cannot be read directly, let alone further applied. Therefore, the research related to image-based table recognition is crucial. It contains the table structure recognition and the table content recognition. Among them, table structure recognition is the most important and difficult task because the table structure is abstract and changeable. In order to address this problem, we propose an innovative table structure recognition method, named TSRDet (Table Structure Recognition based on object Detection). It includes a row-column detection method, named SACNet (StripAttention-CenterNet) and the corresponding post-processing. SACNet is an improved version of the original CenterNet. The specific improvements include the following: firstly, we introduce the Swin Transformer as the encoder to obtain the global feature map of the image. Then, we propose a plug-and-play row-column attention module, including a channel attention module and a row-column spatial attention module. It improves the detection accuracy of rows and columns by capturing long-range row-column feature maps in the image. After completing the row-column detection, this paper also designs a simple and fast post-processing to generate the table structure based on the row-column detection results. Experimental results show that for row-column detection, SACNet has high detection accuracy, even at a high IoU threshold. Specifically, when the threshold is 0.75, its mAP of row detection and column detection still exceeds 90%, which is 91.40% and 92.73% respectively. In addition, in the comparative experiment with the existing object detection methods, SACNet’s performance was significantly better than that of all others. For table structure recognition, the TEDS-Struct score of TSRDet is 95.7%, which shows competitive performance in table structure recognition, and verifies the rationality and superiority of the proposed method.

Extended multi-stream temporal-attention module for skeleton-based human action recognition (HAR)

Graph convolutional networks (GCNs) are an effective skeleton-based human action recognition (HAR) technique. GCNs enable the specification of CNNs to a non-Euclidean frame that is more flexible. The previous GCN-based models still have a lot of issues: (I) The graph structure is the same for all model layers and input data. GCN model's hierarchical structure and human action recognition input diversity make this a problematic approach; (II) Bone length and orientation are understudied due to their significance and variance in HAR. For this purpose, we introduce an Extended Multi-stream Temporal-attention Adaptive GCN (EMS-TAGCN). By training the network topology of the proposed model either consistently or independently according to the input data, this data-based technique makes graphs more flexible and faster to adapt to a new dataset. A spatial, temporal, and channel attention module helps the adaptive graph convolutional layer focus on joints, frames, and features. Hence, a multi-stream framework representing bones, joints, and their motion enhances recognition accuracy. Our proposed model outperforms the NTU RGBD for CS and CV by 0.6% and 1.4%, respectively, while Kinetics-skeleton Top-1 and Top-5 are 1.4% improved, UCF-101 has improved 2.34% accuracy and HMDB-51 dataset has significantly improved 1.8% accuracy. According to the results, our model has performed better than the other models. Our model consistently outperformed other models, and the results were statistically significant that demonstrating the superiority of our model for the task of HAR and its ability to provide the most reliable and accurate results.

Learnable feature alignment with attention-based data augmentation for handling data issue in ancient documents

Recognizing ancient cursive handwritten characters presents unique challenges due to the diversity of writing styles and significant class imbalances, where some characters have disproportionately more samples than others. This imbalance leads to higher misclassification rates for minority classes compared to majority classes. To address these challenges, we propose a novel framework that integrates learnable channel and spatial attention modules to effectively align features between source and target domains for better representation. Our approach incorporates a learnable sequential feature alignment process that dynamically adjusts to the specific characteristics of the data, enhancing the transfer of knowledge across domains. Furthermore, we introduce an attention-based augmentation module to amplify the influence of tail classes. This module leverages class activation maps to identify and augment discriminative features, ensuring the model focuses on the most semantically rich regions, particularly for minority classes. As a result, it aligns the weight norms of minority classes with those of majority classes, effectively mitigating the limitations posed by imbalanced class distributions. This approach effectively mitigates the constraints posed by imbalanced character distributions in ancient handwritten documents. The proposed method increases the accuracy for the CCR, Hanja, Nancho, and Kuzushiji datasets.

Colon polyp detection based on multi-scale and multi-level feature fusion and lightweight convolutional neural network

Early diagnosis and treatment of colorectal polyps are crucial for preventing colorectal cancer. This paper proposes a lightweight convolutional neural network for the automatic detection and auxiliary diagnosis of colorectal polyps. Initially, a 53-layer convolutional backbone network is used, incorporating a spatial pyramid pooling module to achieve feature extraction with different receptive field sizes. Subsequently, a feature pyramid network is employed to perform cross-scale fusion of feature maps from the backbone network. A spatial attention module is utilized to enhance the perception of polyp image boundaries and details. Further, a positional pattern attention module is used to automatically mine and integrate key features across different levels of feature maps, achieving rapid, efficient, and accurate automatic detection of colorectal polyps. The proposed model is evaluated on a clinical dataset, achieving an accuracy of 0.9982, recall of 0.9988, F1 score of 0.9984, and mean average precision (mAP) of 0.9953 at an intersection over union (IOU) threshold of 0.5, with a frame rate of 74 frames per second and a parameter count of 9.08 M. Compared to existing mainstream methods, the proposed method is lightweight, has low operating configuration requirements, high detection speed, and high accuracy, making it a feasible technical method and important tool for the early detection and diagnosis of colorectal cancer.

Convolutional Neural Network Incorporating Multiple Attention Mechanisms for MRI Classification of Lumbar Spinal Stenosis.

Lumbar spinal stenosis (LSS) is a common cause of low back pain, especially in the elderly, and accurate diagnosis is critical for effective treatment. However, manual diagnosis using MRI images is time consuming and subjective, leading to a need for automated methods. This study aims to develop a convolutional neural network (CNN)-based deep learning model integrated with multiple attention mechanisms to improve the accuracy and robustness of LSS classification via MRI images. The proposed model is trained on a standardized MRI dataset sourced from multiple institutions, encompassing various lumbar degenerative conditions. During preprocessing, techniques such as image normalization and data augmentation are employed to enhance the model's performance. The network incorporates a Multi-Headed Self-Attention Module, a Slot Attention Module, and a Channel and Spatial Attention Module, each contributing to better feature extraction and classification. The model achieved 95.2% classification accuracy, 94.7% precision, 94.3% recall, and 94.5% F1 score on the validation set. Ablation experiments confirmed the significant impact of the attention mechanisms in improving the model's classification capabilities. The integration of multiple attention mechanisms enhances the model's ability to accurately classify LSS in MRI images, demonstrating its potential as a tool for automated diagnosis. This study paves the way for future research in applying attention mechanisms to the automated diagnosis of lumbar spinal stenosis and other complex spinal conditions.

Equilibrium Optimization-Based Ensemble CNN Framework for Breast Cancer Multiclass Classification Using Histopathological Image.

Background: Breast cancer is one of the most lethal cancers among women. Early detection and proper treatment reduce mortality rates. Histopathological images provide detailed information for diagnosing and staging breast cancer disease. Methods: The BreakHis dataset, which includes histopathological images, is used in this study. Medical images are prone to problems such as different textural backgrounds and overlapping cell structures, unbalanced class distribution, and insufficiently labeled data. In addition to these, the limitations of deep learning models in overfitting and insufficient feature extraction make it extremely difficult to obtain a high-performance model in this dataset. In this study, 20 state-of-the-art models are trained to diagnose eight types of breast cancer using the fine-tuning method. In addition, a comprehensive experimental study was conducted to determine the most successful new model, with 20 different custom models reported. As a result, we propose a novel model called MultiHisNet. Results: The most effective new model, which included a pointwise convolution layer, residual link, channel, and spatial attention module, achieved 94.69% accuracy in multi-class breast cancer classification. An ensemble model was created with the best-performing transfer learning and custom models obtained in the study, and model weights were determined with an Equilibrium Optimizer. The proposed ensemble model achieved 96.71% accuracy in eight-class breast cancer detection. Conclusions: The results show that the proposed model will support pathologists in successfully diagnosing breast cancer.

A lightweight object detection approach based on edge computing for mining industry

AbstractCoal Mining enterprises deploy numerous monitoring devices to ensure safe and efficient production using target detection technologies. However, deploying deep detection models on edge devices poses challenges due to high computational loads, impacting detection speed and accuracy. A mining target detection dataset has been created to address these issues, featuring key targets in coal mining scenes such as miners, safety helmets, and coal gangue. A model is proposed to improve real‐time performance for edge mining detection tasks. Detection performance is enhanced by incorporating a Pixel‐wise Normalization Spatial Attention Module (PN‐SAM) into the MobileNet‐v3 bneck structure and replacing the h‐swish activation function with Mish, providing more prosperous gradient information transfer. The proposed model, YOLO‐v4‐LSAM, shows a 3.2% mAP improvement on the VOC2012 dataset and a 2.4% improvement on the mining target dataset compared to YOLO‐v4‐Tiny, demonstrating its effectiveness in mining environments. These enhancements enable more accurate and efficient detection in resource‐constrained edge environments, contributing to safer and more reliable monitoring in coal mining operations.

Automated Recognition of Snow-Covered and Icy Road Surfaces Based on T-Net of Mount Tianshan

The Tianshan Expressway plays a crucial role in China’s “Belt and Road” strategy, yet the extreme climate of the Tianshan Mountains poses significant traffic safety risks, hindering local economic development. Efficient detection of hazardous road surface conditions (RSCs) is vital to address these challenges. The complexity and variability of RSCs in the region, exacerbated by harsh weather, make traditional surveillance methods inadequate for real-time monitoring. To overcome these limitations, a vision-based artificial intelligence approach is urgently needed to ensure effective, real-time detection of dangerous RSCs in the Tianshan road network. This paper analyzes the primary structures and architectures of mainstream neural networks and explores their performance for RSC recognition through a comprehensive set of experiments, filling a research gap. Additionally, T-Net, specifically designed for the Tianshan Expressway engineering project, is built upon the optimal architecture identified in this study. Leveraging the split-transform-merge structure paradigm and asymmetric convolution, the model excels in capturing detailed information by learning features across multiple dimensions and perspectives. Furthermore, the integration of channel, spatial, and multi-head attention modules enhances the weighting of key features, making the T-Net particularly effective in recognizing the characteristics of snow-covered and icy road surfaces. All models presented in this paper were trained on a custom RSC dataset, compiled from various sources. Experimental results indicate that the T-Net outperforms fourteen once state-of-the-art (SOTA) models and three models specifically designed for RSC recognition, with 97.44% accuracy and 9.79% loss on the validation set.

Retinal fundus image super-resolution based on generative adversarial network guided with vascular structure prior

Many ophthalmic and systemic diseases can be screened by analyzing retinal fundus images. The clarity and resolution of retinal fundus images directly determine the effectiveness of clinical diagnosis. Deep learning methods based on generative adversarial networks are used in various research fields due to their powerful generative capabilities, especially image super-resolution. Although Real-ESRGAN is a recently proposed method that excels in processing real-world degraded images, it suffers from structural distortions when super-resolving retinal fundus images are rich in structural information. To address this shortcoming, we first process the input image using a pre-trained U-Net model to obtain a structural segmentation map of the retinal vessels and use the segmentation map as the structural prior. The spatial feature transform layer is then used to better integrate the structural prior into the generation process of the generator. In addition, we introduce channel and spatial attention modules into the skip connections of the discriminator to emphasize meaningful features and accordingly enhance the discriminative power of the discriminator. Based on the original loss functions, we introduce the L1 loss function to measure the pixel-level differences between the segmentation maps of retinal vascular structures in the high-resolution images and the super-resolution images to further constrain the super-resolution images. Simulation results on retinal image datasets show that our improved algorithm results have a better visual performance by suppressing structural distortions in the super-resolution images.

ResMT: A hybrid CNN-transformer framework for glioma grading with 3D MRI

Accurate grading of gliomas is crucial for treatment strategies and prognosis. While convolutional neural networks (CNNs) have proven effective in classifying medical images, they struggle with capturing long-range dependencies among pixels. Transformer-based networks can address this issue, but CNN-based methods often perform better when trained on small datasets. Additionally, tumor segmentation is essential for classification models, but training an additional segmentation model significantly increases workload. To address these challenges, we propose ResMT, which combines CNN and transformer architectures for glioma grading, extracting both local and global features efficiently. Specifically, we designed a spatial residual module (SRM) where a 3D CNN captures glioma's volumetric complexity, and Swin UNETR, a pre-trained segmentation model, enhances the network without extra training. Our model also includes a multi-plane channel and spatial attention module (MCSA) to refine the analysis by focusing on critical features across multiple planes (axial, coronal, and sagittal). Transformer blocks establish long-range relationships among planes and slices. We evaluated ResMT on the BraTs19 dataset, comparing it with baselines and state-of-the-art models. Results demonstrate that ResMT achieves the highest prediction performance with an AUC of 0.9953, highlighting hybrid CNN-transformer models' potential for 3D MRI classification.

Spatial attention for human-centric visual understanding: An Information Bottleneck method

The selective visual attention mechanism in the Human Visual System (HVS) restricts the amount of information that reaches human visual awareness, allowing the brain to perceive high-fidelity natural scenes in real-time with limited computational cost. This selectivity acts as an “Information Bottleneck (IB)” that balances information compression and predictive accuracy. However, such information constraints are rarely explored in the attention mechanism for deep neural networks (DNNs). This paper introduces an IB-inspired spatial attention module for DNNs, which generates an attention map by minimizing the mutual information (MI) between the attentive content and the input while maximizing that between the attentive content and the output. We develop this IB-inspired attention mechanism based on a novel graphical model and explore various implementations of the framework. We show that our approach can yield attention maps that neatly highlight the regions of interest while suppressing the backgrounds, and are interpretable for the decision-making of the DNNs. To validate the effectiveness of the proposed IB-inspired attention mechanism, we apply it to various computer vision tasks including image classification, fine-grained recognition, cross-domain classification, semantic segmentation, and object detection. Extensive experiments demonstrate that it bootstraps standard DNN structures quantitatively and qualitatively for these tasks.

Semi-supervised recognition for artificial intelligence assisted pathology image diagnosis

The analysis and interpretation of cytopathological images are crucial in modern medical diagnostics. However, manually locating and identifying relevant cells from the vast amount of image data can be a daunting task. This challenge is particularly pronounced in developing countries where there may be a shortage of medical expertise to handle such tasks. The challenge of acquiring large amounts of high-quality labelled data remains, many researchers have begun to use semi-supervised learning methods to learn from unlabeled data. Although current semi-supervised learning models partially solve the issue of limited labelled data, they are inefficient in exploiting unlabeled samples. To address this, we introduce a new AI-assisted semi-supervised scheme, the Reliable-Unlabeled Semi-Supervised Segmentation (RU3S) model. This model integrates the ResUNet-SE-ASPP-Attention (RSAA) model, which includes the Squeeze-and-Excitation (SE) network, Atrous Spatial Pyramid Pooling (ASPP) structure, Attention module, and ResUNet architecture. Our model leverages unlabeled data effectively, improving accuracy significantly. A novel confidence filtering strategy is introduced to make better use of unlabeled samples, addressing the scarcity of labelled data. Experimental results show a 2.0% improvement in mIoU accuracy over the current state-of-the-art semi-supervised segmentation model ST, demonstrating our approach’s effectiveness in solving this medical problem.

Target Recognition Based on Infrared and Visible Image Fusion and Improved YOLOv8 Algorithm.

In response to the issue that the fusion process of infrared and visible images is easily affected by lighting factors, in this paper, we propose an adaptive illumination perception fusion mechanism, which was integrated into an infrared and visible image fusion network. Spatial attention mechanisms were applied to both infrared images and visible images for feature extraction. Deep convolutional neural networks were utilized for further feature information extraction. The adaptive illumination perception fusion mechanism is then integrated into the image reconstruction process to reduce the impact of lighting variations in the fused images. A Median Strengthening Channel and Spatial Attention Module (MSCS) was designed to be integrated into the backbone of YOLOv8. In this paper, we used the fusion network to create a dataset named ivifdata for training the target recognition network. The experimental results indicated that the improved YOLOv8 network saw further enhancements of 2.3%, 1.4%, and 8.2% in the Recall, mAP50, and mAP50-95 metrics, respectively. The experiments revealed that the improved YOLOv8 network has advantages in terms of recognition rate and completeness, while also reducing the rates of false negatives and false positives.

Music source feature extraction based on improved attention mechanism and phase feature

Music source feature extraction is an important research direction in music information retrieval and music recommendation system. To extract the features of music sources more effectively, the study introduces the jump attention mechanism and combines it with the convolutional attention module. Also, a feature extraction module based on Unet + + and spatial attention module is proposed. In addition, the phase feature information of the mixed music signals is utilized to improve the network performance. Results showed that this model was studied to perform well in music source separation experiments of vocals and accompaniment. For vocal separation on the MIR-1K dataset, the model achieves 11.25 dB, 17.34 dB, and 13.83 dB for each metric, respectively. Meanwhile, for drum separation on the DSD100 dataset, the model achieves a median signal-to-source distortion ratio of 4.36 dB, which is 2.91 dB better than that of the Spectral Hierarchical Network model. For the separation of the bass sound and the human voice, the model's in the separation of bass and human voice, the median distortion ratio of the model is as high as 4.87 dB and 6.09 dB, which is better than that of the Spectral Hierarchical Network model. This indicates the significant performance advantages in feature extraction and separation of music sources, and it has important application values in music production and speech recognition.