The Classification and Recognition of Wrestling Injury Based on Deep Learning

This study proposes a deep learning (DL)-based multi-class injury classification and recognition model to improve the efficiency and accuracy of injury diagnosis in wrestling. It also optimizes the design by combining Convolutional Neural Networks (CNNs), residual networks, and dense connection networks. Since the existing diagnostic methods rely on manual experience and the classification is not fine enough, this study designs a model framework to automatically classify fractures, sprains, strains, and other common wrestling injuries. The main problems include improving the model’s classification accuracy, enhancing the reasoning efficiency, optimizing the computational complexity, and solving data imbalance and poor classification of small sample categories. By comparing advanced models such as EfficientNet, Vision Transformer (ViT), and Swin Transformer, this study comprehensively evaluates the model performance in many dimensions. It encompasses classification accuracy, specificity, Area Under the Receiver Operating Characteristic Curve (AUC-ROC), and average cross-entropy loss. Regarding classification accuracy, the proposed optimized model’s performance in competition, training, and accidental injuries is 0.907, 0.930, and 0.890, respectively, higher than most comparative models. In terms of specificity and AUC-ROC, the proposed model’s AUC-ROC value in training injury reaches 0.958, showing excellent classification performance. In addition, the average cross-entropy loss is as low as 0.242 in competition injury, showing that the model can maintain a low classification error under multiple injury types. Therefore, this study has made an important contribution to injury classification and recognition technology based on DL in the sports medicine field. Especially, in improving classification efficiency and accuracy, it provides reference and support for applying intelligent medical technology in the future.