Abstract: Aiming at the problem of complex image content, diverse action postures, and significant background interference in the task of sports image classification,  we proposed a sports image classification algorithm based on an improved ResNet50 model. Firstly,  a squeeze-and-excitation module was embedded within the residual structure to adaptively enhance key channel features and improve feature expression capability. Secondly, on this basis, a feature pyramid network was introduced to achieve effective fusion of multi-scale features, and enhance the model’s perception ability of objects at different scales. Finally, classification prediction was performed through global average pooling and  fully connected layers. Experimental results show  that the classification accuracy of the proposed method is about  5% higher than that of the conventional ResNet50 model, fully demonstrating  its robustness and superiority in handling complex actions and diverse backgrounds. The experimental results  not only validate the effectiveness and feasibility of the proposed method,  but also provide more reliable technical support and practical reference for applications in sports video analysis,  intelligent sports training and other related fields.

Key words: deep residual network, sports image classification, ResNet50 model, attention mechanism, mult-scale feature fusion