A motion injury image classification method and system based on deep learning

Abstract

This proposal discloses a deep learning-based motion injury image classification method and system. Based on the traditional ResNet-50 network, the input module is replaced with the downsampling feature extraction module proposed in this proposal, and the original bottleneck structure in the final stage of the network is replaced with the lightweight multi-scale feature fusion and enhancement module proposed in this proposal. The original bottleneck structure in the intermediate stage is retained, realizing multi-scale feature extraction and fusion of motion injury images. While effectively controlling the number of model parameters and computational complexity, it significantly improves the recognition accuracy and stability of various types of motion injuries.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a motion injury image classification method and system based on deep learning. Background Art

[0002] Traditionally, medical professionals rely primarily on clinical manifestations and imaging data for sports injury diagnosis and treatment. Conventional methods typically involve manual review of X-rays, magnetic resonance imaging (MRI), or ultrasound images to determine whether injuries to tissues such as ligaments, bones, or muscles are present. This diagnostic approach relies heavily on specialized skills, and subjective factors can also impact diagnostic efficiency and accuracy. With the rapid development of artificial intelligence, AI-based sports injury recognition methods have emerged. However, some traditional image processing and pattern recognition algorithms remain deficient in recognition performance and feature extraction depth, making them ineffective against complex sports injury images. Early deep learning models suffer from large parameter counts, high computational costs, and insufficient multi-scale feature extraction capabilities. With the continuous evolution of deep learning technology, improvements to network architectures and the integration of multi-scale feature extraction and attention mechanisms have significantly improved the accuracy of sports injury image classification. However, balancing network optimization with practical application requirements remains a challenge. Reducing network parameters and computational complexity while maintaining classification accuracy, while further enhancing the ability to extract and represent multi-scale features, has become a pressing technical challenge in this field. Based on this, this paper proposes a motion injury image classification method based on deep learning, aiming to achieve more accurate and efficient classification in complex motion injury image recognition scenarios. Summary of the Invention

[0003] The purpose of this invention is to address the above-mentioned issues and propose a deep learning-based motion injury image classification method and system. The input backbone of the basic ResNet-50 structure is replaced with a downsampling feature extraction module (SCR-DS-Block), and the last bottleneck structure (×3) of the residual block is replaced with a lightweight multi-scale feature fusion and enhancement module (EMA-Bottleneck). The original bottleneck structure of the intermediate stage is retained, enabling multi-scale feature extraction and fusion of motion injury images. While effectively controlling the number of model parameters and computational complexity, the recognition accuracy and stability of various types of motion injuries are significantly improved.

[0004] A motion injury image classification method based on deep learning includes constructing a classification network including an input layer, a residual block, an average pooling layer and a fully connected layer.

[0005] The input layer includes a downsampling feature extraction module based on depthwise separable convolution and SCConv (Spatial and Channel Reconstruction Convolution) optimization;

[0006] The residual block includes several bottleneck structures and a lightweight multi-scale feature fusion and enhancement module;

[0007] The lightweight multi-scale feature fusion and enhancement module includes depthwise separable convolution and dual-path attention mechanism;

[0008] The classification network is trained using a pre-prepared image dataset containing motion injury images to obtain a motion injury image classification model that can determine whether the motion injury image is injured. The motion injury image classification labels include two category labels: normal and injury.

[0009] In the above-mentioned deep learning-based motion injury image classification method, the downsampling feature extraction module includes an initial layer convolution, an SCConv module, batch normalization, an activation function and medium- and large-size depthwise separable convolution, such as a 7×7 depthwise separable convolution, which are connected in sequence to simultaneously complete downsampling and efficient feature extraction.

[0010] In the above-mentioned motion injury image classification method based on deep learning, the residual block includes a three-layer bottleneck structure, wherein the first layer includes three bottleneck structures, the second layer includes four bottleneck structures, and the third layer includes six bottleneck structures;

[0011] The feature maps initially extracted by the downsampling feature extraction module are input into the three bottleneck structures of the first layer;

[0012] The feature maps processed by the six bottleneck structures in the third layer are input into the lightweight multi-scale feature fusion and enhancement module;

[0013] The residual block includes three lightweight multi-scale feature fusion and enhancement modules arranged in parallel.

[0014] In the above-mentioned motion injury image classification method based on deep learning, the feature map processed by the bottleneck structure is input into the lightweight multi-scale feature fusion and enhancement module;

[0015] The output of the lightweight multi-scale feature fusion and enhancement module is connected to the average pooling layer;

[0016] The output of the average pooling layer is connected to the fully connected layer.

[0017] In the above-mentioned motion injury image classification method based on deep learning, the lightweight multi-scale feature fusion and enhancement module includes:

[0018] Dimensionality reduction convolution layer, used to reduce the channel dimension of the input feature map to reduce the amount of calculation;

[0019] A lightweight EMA attention mechanism is used to perform spatial and channel-wise attention operations on the feature maps after dimensionality reduction, enhancing the representation capability of features.

[0020] The dimension-increasing convolution layer is used to increase the channel dimension of the feature map processed by the attention mechanism and restore it to the original number of channels;

[0021] The input of the dimensionality reduction convolution layer and the output of the dimensionality increase convolution layer are residually connected.

[0022] In the above-mentioned motion injury image classification method based on deep learning, the dimensionality reduction convolution layer includes a 1×1 dimensionality reduction convolution;

[0023] The dimensionality-increasing convolution layer includes a 1×1 dimensionality-increasing convolution.

[0024] In the above-mentioned motion injury image classification method based on deep learning, the feature map output by the dimensionality reduction convolution layer is processed by the batch normalization layer and then input into the lightweight EMA attention mechanism;

[0025] The feature map processed by the lightweight EMA attention mechanism is processed by the batch normalization layer and the activation function layer and then input into the dimension-raising convolution layer described above;

[0026] The feature map output by the residual connection is processed by the activation function layer and then input into the average pooling layer.

[0027] In the above-mentioned deep learning-based motion injury image classification method, the lightweight EMA attention mechanism includes the following steps:

[0028] Perform small-size separable depth convolution operations on the feature map to further extract features;

[0029] Perform batch normalization and activation function processing on the feature map after further feature extraction;

[0030] Perform spatial attention on the output of batch normalization and activation function processing, including horizontal / vertical adaptive pooling, 1×1 convolution and activation function processing, to highlight important spatial areas in the feature map;

[0031] The feature map output by the spatial attention operation is multiplied and fused with the feature map output by batch normalization and activation function processing to output spatial enhanced features;

[0032] Perform channel attention operations on the spatial enhancement features, including global average pooling, lightweight MLP (multi-layer perceptron) and Sigmoid activation function processing, to enhance the features of important channels. The Sigmoid activation function is a nonlinear activation function.

[0033] Multiplying the spatial enhancement feature with the output of the channel attention sequence to output an enhanced feature map;

[0034] The enhanced features are sent to the average pooling layer for processing.

[0035] In the above-mentioned deep learning-based motion injury image classification method, before using the image dataset to train the classification network, it also includes pre-training the classification network using a dataset containing a large number of skin lesion images.

[0036] A motion injury image classification system based on deep learning includes a motion injury image classification model constructed by the above method, which is used to receive an image to be predicted and output a judgment result of whether there is an injury based on the image to be predicted.

[0037] The advantages of the present invention are:

[0038] This scheme is based on the traditional ResNet-50 network, replacing the input module with the downsampling feature extraction module proposed in this scheme, and replacing the original bottleneck structure (×3) in the final stage of the network with the lightweight multi-scale feature fusion and enhancement module proposed in this scheme. It adopts a balanced design strategy, which not only significantly enhances the multi-scale feature extraction capability of motion damage images, but also makes more efficient use of computing resources by selectively applying the attention mechanism.

[0039] This solution applies complex attention mechanisms only to the high-level semantic feature processing stage, ensuring the enhancement of key features while avoiding excessive computational overhead in the early feature extraction stages. The downsampling feature extraction module replaces standard convolution with depthwise separable convolution, reducing the number of parameters by approximately 9 times compared to traditional convolution. The lightweight multi-scale feature fusion and enhancement module uses a dual-path attention design, both spatially and channel-wise, to avoid the complex grouping operations and matrix multiplications required in the traditional EMA module, further reducing computational overhead.

[0040] Based on the above design, this solution method achieves the effect of lower computational complexity and parameter number than the original ResNet-50 model while maintaining high classification accuracy.

[0041] By adopting differentiated hyperparameters and optimization strategies in the pre-training and formal training stages, this solution can achieve rapid convergence and efficient expansion in image scenes of different scales and types, thereby better balancing accuracy and efficiency in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the method for classifying motion injury images based on deep learning of the present invention;

[0043] Figure 2 This is a model structure diagram of the motion injury image classification network of the motion injury image classification method based on deep learning of the present invention;

[0044] Figure 3 This is a module structure diagram of the SCR-DS-Block in the motion injury image classification method based on deep learning of the present invention;

[0045] Figure 4 This is a module structure diagram of the lightweight EMA-Bottleneck in the deep learning-based motion injury image classification method of the present invention;

[0046] Figure 5 The confusion matrix of the proposed method and five other classic deep learning models on the motion injury image classification task;

[0047] Figure 6 Comparison chart of loss curve and FLOPs efficiency curve of the provided embodiment;

[0048] Figure 7 Comparison of the accuracy and F1-Score curves of each model during training;

[0049] Figure 8 This is an example diagram of the sports injury dataset. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1As shown, this embodiment provides a motion injury image classification method based on deep learning, which can not only complete the motion injury classification task more efficiently but also improve the classification accuracy and precision. First, data preprocessing is performed, and the motion injury image dataset after data preprocessing is divided into a training set, a validation set, and a test set; then a motion injury image classification network is constructed, and the ImageNet image dataset is used to pre-train the motion injury image classification network to obtain a pre-trained image classification network, and then the training parameters are adjusted, and the motion injury training set and validation set are input into the constructed and pre-trained classification network for training to obtain a trained motion injury image classification model. The test set is input into the motion injury image classification model to obtain the results of the motion injury image classification to verify the performance of the model. The specific implementation steps are as follows:

[0052] Step 1: Obtain a sports injury image dataset and its corresponding classification labels, preprocess and analyze the data, and then divide the preprocessed sports injury dataset into a training set, a validation set, and a test set according to a certain ratio.

[0053] Step 1.1: Resize each image in the motion injury image dataset to 224 pixels * 224 pixels;

[0054] Step 1.2: Perform various data augmentation operations such as random rotation, random cropping, random flipping, and random scaling on the motion injury image dataset after the image size adjustment in step 1.1;

[0055] Step 1.3: After the data augmentation operation in step 1.2, the motion injury image dataset is divided into a training set, a validation set, and a test set according to the proportion, of which 60% is the training set, 20% is the validation set, and 20% is the test set. The motion injury image classification labels include two categories: normal and injury.

[0056] Step 2: Construct a motion injury image classification network. The structure of this motion injury image classification network is as follows: Figure 2 As shown in the figure, the network is optimized and improved based on the ResNet-50 structure. The original input backbone is replaced by a downsampling feature extraction module (SCR-DS-Block), and the original Bottleneck structure (×3, ×4, ×6) in the middle stage is retained. Only in the last stage of the network is the bottleneck structure (×3) replaced by a lightweight multi-scale feature fusion and enhancement module (EMA-Bottleneck).

[0057] like Figure 3As shown in the figure, the SCR-DS-Block module consists of a 3×3 initial convolution (stride = 2) → SCConv (Spatial and Channel Reconstruction Convolution) → batch normalization (BatchNorm) + ReLU activation function → 7×7 depth-wise separable convolution (stride = 2), which can simultaneously achieve downsampling and efficient feature extraction.

[0058] After initial processing by the SCR-DS-Block module, the backbone feature map is further passed to the lightweight EMA-Bottleneck module for in-depth multi-scale feature extraction. This module not only retains residual connections to enhance gradient flow stability but also utilizes a lightweight EMA attention mechanism to efficiently capture and fuse critical multi-scale information, while significantly reducing computational costs. After feature extraction by the multi-layer EMA-Bottleneck module, the network generates a high-quality feature representation map, which is then processed through average pooling and fully connected layers to output the final classification result.

[0059] like Figure 4 As shown, the lightweight EMA-Bottleneck module consists of: 1×1 dimensionality reduction convolution → lightweight EMA attention mechanism → 1×1 dimensionality increase convolution, while retaining residual connections. The EMA mechanism first performs a 3×3 depthwise convolution on the features, then applies spatial attention (horizontally / vertically adaptive pooling + 1×1 convolution + sigmoid) and channel attention (global average pooling + lightweight MLP + sigmoid activation function) sequentially. These two stages are then fused through sequential multiplication to output enhanced features. This design not only retains the feature enhancement capabilities of the attention mechanism but also significantly reduces computational complexity and parameter count. The lightweight EMA-Bottleneck module not only retains residual connections to enhance gradient flow stability but also uses a lightweight attention mechanism to efficiently capture and fuse critical multi-scale information. After feature extraction by the multi-layer EMA-Bottleneck module, the network generates a high-quality feature representation graph, which, through average pooling and fully connected layers, outputs the final classification result.

[0060] Step 3: Use the ISIC2019 dataset to pre-train the constructed sports injury image classification network to obtain a pre-trained classification model. Specifically:

[0061] Step 3.1: Initialize the parameters of the motion injury image classification network;

[0062] Step 3.2: Set the training parameters, the number of training rounds is 200, the sample batch size is 16, the network initialization learning rate is set to 1e-3, and the Adam optimizer is used for optimization;

[0063] Step 3.3: Use ACC (Accuracy) as the model evaluation criterion to modify the training parameters of the classification model;

[0064] Step 3.4: Input the training set of the ISIC2019 image dataset into the constructed motion injury image classification network for iterative training, ultimately obtaining a pre-trained image classification model. ISIC2019 is a large dataset containing 25,331 skin lesion images. Its diversity and complexity help the network learn universal feature representation capabilities, providing good parameter initialization for subsequent motion injury recognition tasks.

[0065] Step 4: Input the training set and validation set into the motion injury classification network for model training to obtain a trained motion injury image classification model. Specifically:

[0066] Step 4.1: The network parameters for motion injury image classification use the weights pre-trained in step 3;

[0067] Step 4.2: Set the training parameters, the number of training rounds is 100, the sample batch size is 32, the network initialization learning rate is set to 1e-4, and the Adam optimizer is used for optimization;

[0068] Step 4.3: Use ACC (Accuracy) as the model evaluation criterion to modify the training parameters of the classification model;

[0069] Step 4.4: Input the training set of the motion injury image dataset into the constructed motion injury image classification network for iterative training, and finally obtain a trained image classification model.

[0070] Step 5: Input the test set of the sports injury image dataset into the trained classification model, obtain the classification results and calculate the accuracy. Through comprehensive analysis with other evaluation indicators, verify the generalization ability and classification performance of the model on unseen data.

[0071] The proposed solution addresses existing issues with motion injury image classification methods, including feature extraction depth, network parameter redundancy, insufficient multi-scale feature representation, and difficulty balancing accuracy and efficiency in practical applications. By replacing the input backbone of the basic ResNet-50 architecture with an SCR-DS-Block module and replacing the original bottleneck structure (×3) in the final stage of the network with a lightweight EMA-Bottleneck module, while retaining the original Bottleneck structure in the intermediate stages, this approach achieves multi-scale feature extraction and fusion for motion injury images. While effectively controlling the number of model parameters and computational complexity, it significantly improves the accuracy and stability of recognition for various types of motion injuries. The lightweight EMA-Bottleneck module innovatively combines depthwise separable convolution with a dual-path attention mechanism, significantly reducing computational cost while maintaining feature enhancement capabilities. This solution also sets different hyperparameters and optimization strategies for both pre-training and training, further enhancing the network's adaptability to small sample sizes, imbalanced data, and noisy environments. The overall system exhibits excellent scalability and robustness.

[0072] To verify the effectiveness of the method proposed in this solution, this example evaluated the performance of the proposed method on a motion injury image dataset. The motion injury image dataset is derived from images of injured parts of patients with sports injuries collected by multiple hospitals and medical institutions. The dataset contains categories of normal and injury, and is divided into training set, validation set, and test set, with a ratio of 6:2:2 respectively.

[0073] This method was developed in PyTorch and trained on three 3090 GPUs. During training, the Adam optimizer was used to optimize the parameters of the denoising network. The batch size was set to 32, the initial learning rate was set to 1e-4, and a total of 100 training rounds were performed.

[0074] In this experiment, we used accuracy (ACC) and F1-Score as core metrics to evaluate the model's classification performance. Accuracy measures the model's classification accuracy across all samples, as defined in Formula (1). As a widely used evaluation metric, accuracy can intuitively reflect the model's overall classification performance.

[0075] (1)

[0076] In addition, the F1-Score is used to further evaluate the model's balance between precision and recall. The F1-Score is the harmonic mean of precision and recall, and is defined as shown in Formula (2). In the evaluation of classification models, precision and recall are used to measure the model's accuracy and recall in identifying positive samples, respectively. Precision (P) represents the proportion of true positive examples among the positive samples identified by the model, while recall (R) represents the proportion of correctly identified positive examples among all positive samples.

[0077] (2)

[0078] To calculate the F1-Score, experiments counted four categories of samples: TP (True Positives), TN (True Negatives), FP (False Positives), and FN (False Negatives). The F1-Score effectively reflects a model's classification performance on imbalanced datasets by comprehensively considering the balance between precision and recall. Generally, a higher F1-Score indicates a better balance between the model's ability to accurately identify positive and negative samples, and thus, better classification performance. Therefore, in tasks with an imbalanced class distribution, the F1-Score can be more valuable than accuracy.

[0079] In the evaluation and application of deep learning models, it is important not only to focus on classification performance metrics (such as accuracy and F1-Score), but also to consider the model's computational overhead and scale. To this end, this experiment introduces floating-point operations (FLOPs) and parameter count (Param) as important reference indicators for measuring model complexity. FLOPs is used to measure the total number of floating-point operations required by the model when performing a forward propagation (or inference). As a common metric for measuring computational complexity, a larger FLOPs value generally indicates a higher network depth and structural complexity of the model, which in turn increases the hardware resources and time cost required for inference. By calculating FLOPs under different structures, we can intuitively compare the differences in computational load between models during actual deployment, thus providing a basis for lightweight improvements to the model.

[0080] Meanwhile, the number of parameters (Param) measures the scale of the model. A higher number of parameters often indicates a model's greater potential in learning and expressing complex features, but this also leads to higher storage requirements and training costs. For certain application scenarios with limited hardware resources or high latency requirements, keeping the number of parameters within a reasonable range is key to maintaining a balance between model efficiency and performance.

[0081] In this experiment, we compared the computational complexity and scale of various methods by statistically analyzing the FLOPs and Param values ​​of different models under different network structures. Combined with classification metrics, these results help us comprehensively evaluate the trade-off between accuracy and efficiency, providing a reliable reference for selecting appropriate network structures in practical application scenarios.

[0082] Table 1. Comparison of objective evaluation of different methods on motion injury image dataset

[0083]

[0084] The experimental results in Table 1 show that our proposed method achieves state-of-the-art performance in classification metrics such as accuracy (ACC) and F1-score on the motion injury image dataset. Furthermore, in terms of FLOPs (3.7G) and parameter count (24.3M), our proposed method demonstrates ideal computational cost and model size, with both computational complexity and parameter count lower than the original ResNet-50 (4.1G FLOPs, 25.6M Param). This is primarily due to the use of depthwise separable convolutions in the SCR-DS-Block module and the efficient attention design in the lightweight EMA-Bottleneck module. Compared to other mainstream models, our proposed method also demonstrates improved operational efficiency while maintaining high classification accuracy.

[0085] To further verify the superiority of this solution, this embodiment provides visualization results of the confusion matrix, training process curve, and performance efficiency comparison.

[0086] Figure 5 The confusion matrix of this method and five other classic deep learning models on the motion injury image classification task is presented. The confusion matrix clearly shows that this method has a significant advantage in classification results. This method correctly identified 498 cases of injury categories, significantly higher than comparison methods such as ResNet-50 (477 cases), EfficientNet-B5 (479 cases), and MobileNet-V2 (449 cases). At the same time, in the correct identification of the normal category, this method also outperformed other models with 732 cases. More importantly, this method had the lowest number of false positives (48 cases) and false negatives (22 cases) among all comparison methods, which is of great significance for medical diagnostic applications and fully demonstrates its significant effect in reducing the risk of misdiagnosis and missed diagnosis.

[0087] Figure 6 A comparison of the loss curve and the FLOPs efficiency curve is provided.

[0088] The loss curve shows that the training loss of this method (red) decreases the fastest and has the lowest final convergence value (about 0.28), which is significantly better than other comparison models, indicating that it has better optimization effect and learning ability.

[0089] The FLOPs efficiency curve shows the accuracy-to-computational complexity ratio (ACC / FLOPs) of each model, representing classification performance per unit of computational effort. Our approach outperforms all other compared methods on this metric, demonstrating its superior resource efficiency. While lightweight networks like MobileNet-V2 have extremely low FLOPs, their accuracy is also relatively low. Our approach, through innovative network structure design, achieves higher classification accuracy while maintaining low computational complexity.

[0090] Figure 7 The accuracy curve and F1-Score curve of each model during training are shown. From the curve graph, we can observe:

[0091] The proposed method (red curve) maintains the highest accuracy and F1-Score value throughout the training process, and the upward trend is more stable, indicating that the learning ability and generalization performance of this model are better than those of the comparison method.

[0092] In the early stages of training (about the first 20 rounds), the performance of this approach improves significantly faster than that of other models, reflecting the positive impact of the SCR-DS-Block and lightweight EMA-Bottleneck modules on the model convergence speed.

[0093] In the later stages of training (80-100 rounds), the accuracy of our method stabilized at around 0.94, and the F1-Score also reached a similar level. The corresponding indicators of the closest comparison method, EfficientNet-B5, were lower, demonstrating the advantage of our method in sports injury recognition accuracy.

[0094] The above charts and graphs show that our proposed method not only outperforms existing methods in classification accuracy for motion injury image classification, but also excels in training convergence speed, computational efficiency, and reduced misdiagnosis risk. These results fully demonstrate the effectiveness of the SCR-DS-Block module and the lightweight EMA-Bottleneck module in feature extraction and multi-scale information fusion, as well as the potential value of this approach in practical medical image analysis applications.

[0095] The specific embodiments described in this embodiment are merely examples of the spirit of this solution. Those skilled in the art of the present solution may make various modifications or additions to the described specific embodiments or replace them with similar methods without departing from the spirit of this solution or exceeding the scope defined by the appended claims.

Claims

1. A motion injury image classification method based on deep learning, comprising constructing a classification network comprising an input layer, a residual block, an average pooling layer, and a fully connected layer, characterized in that: The input layer includes a downsampling feature extraction module based on depthwise separable convolution and SCConv optimization; The residual block includes several bottleneck structures and a lightweight multi-scale feature fusion and enhancement module; The lightweight multi-scale feature fusion and enhancement module includes depthwise separable convolution and dual-path attention mechanism; Using a pre-prepared image dataset containing motion injury images to train the classification network to obtain a motion injury image classification model capable of determining whether the motion injury images are injured; The downsampling feature extraction module includes an initial layer convolution, an SCConv module, a batch normalization, an activation function and a medium and large size depthwise separable convolution connected in sequence; The feature map processed by the bottleneck structure is input into the lightweight multi-scale feature fusion and enhancement module; The output of the lightweight multi-scale feature fusion and enhancement module is connected to the average pooling layer; The output of the average pooling layer is connected to the fully connected layer; The lightweight multi-scale feature fusion and enhancement module includes: Dimensionality reduction convolution layer, used to reduce the channel dimension of the input feature map; A lightweight EMA attention mechanism is used to perform spatial and channel-wise attention operations on the feature maps after dimensionality reduction. The dimension-increasing convolution layer is used to increase the channel dimension of the feature map processed by the attention mechanism and restore it to the original number of channels; The input of the dimensionality reduction convolution layer and the output of the dimensionality increase convolution layer are residually connected; The lightweight EMA attention mechanism includes the following steps: Perform small-size separable depth convolution operations on the feature map to further extract features; Perform batch normalization and activation function processing on the feature map after further feature extraction; Perform spatial attention on the batch normalized and activated output, including horizontal / vertical adaptive pooling, 1×1 convolution, and activation. The feature map output by the spatial attention operation is multiplied and fused with the feature map output by batch normalization and activation function processing to output spatial enhanced features; Perform channel attention operations on the spatial enhancement features, including global average pooling, lightweight MLP and Sigmoid activation function processing; Multiplying the spatial enhancement feature with the output of the channel attention sequence to output an enhanced feature map; The enhanced features are sent to the average pooling layer for processing.

2. The motion injury image classification method based on deep learning according to claim 1, characterized in that: The residual block includes a three-layer bottleneck structure, wherein the first layer includes three bottleneck structures, the second layer includes four bottleneck structures, and the third layer includes six bottleneck structures; The feature maps initially extracted by the downsampling feature extraction module are input into the three bottleneck structures of the first layer; The feature maps processed by the six bottleneck structures in the third layer are input into the lightweight multi-scale feature fusion and enhancement module; The residual block includes three lightweight multi-scale feature fusion and enhancement modules arranged in parallel.

3. The motion injury image classification method based on deep learning according to claim 1, characterized in that: The dimensionality reduction convolution layer includes a 1×1 dimensionality reduction convolution; The dimensionality-increasing convolution layer includes a 1×1 dimensionality-increasing convolution.

4. The method for classifying motion injury images based on deep learning according to claim 1, characterized in that: The feature map output by the dimensionality reduction convolution layer is processed by the batch normalization layer and then input into the lightweight EMA attention mechanism; The feature map processed by the lightweight EMA attention mechanism is processed by the batch normalization layer and the activation function layer and then input into the dimension-raising convolution layer described above; The feature map output by the residual connection is processed by the activation function layer and then input into the average pooling layer.

5. The method for classifying motion injury images based on deep learning according to claim 1, characterized in that: Before using the image dataset to train the classification network, the method also includes pre-training the classification network using a dataset containing a large number of skin lesion images.

6. A deep learning-based sports injury image classification system, characterized by: A motion injury image classification model constructed by the method according to any one of claims 1 to 5 is used to receive an image to be predicted and output a judgment result of whether the image to be predicted is injured based on the image to be predicted.

Images