Dynamic scale fusion and cascaded prototype refinement for few-shot sports image classification

Inroduction Sports image classification plays a crucial role in intelligent sports analytics by supporting athlete behavior analysis, training feedback, exercise participation monitoring, and data-driven sports behavior studies. Although deep learning methods have advanced sports image recognition by learning hierarchical visual patterns from image data, they usually require abundant annotations, which limits their applicability in few-shot scenarios. In addition, sports images often show large-scale variations, severe background distraction, and fine-grained inter-class similarities, further degrading recognition performance under scarce supervision. Methods To address these challenges, this study proposes a dynamic scale fusion and cascaded prototype refinement network (DSCPNet) for few-shot sports image classification. First, Dynamic Scale Fusion Attention (DSFA) is introduced to dynamically select and fuse discriminative contextual cues under varying receptive fields, thereby enhancing scale-adaptive feature representation from scarce support samples. Second, Cascaded Foreground-Scene Interaction (CFSI) is designed to suppress background interference through cascaded interaction modulation, strengthening target-focused embeddings and reducing intra-class variance. Finally, Cross-Prototype Feature Refinement (CPFR) is developed to calibrate support prototypes and query representations in a cross-prototype manner, producing more discriminative class prototypes and improving the robustness of similarity estimation for fine-grained recognition. Results Experiments on two sports-oriented few-shot benchmark datasets show that DSCPNet consistently achieves stronger recognition performance than competing methods in both 1-shot and 5-shot scenarios. Discussion These results indicate that DSCPNet can improve few-shot sports image recognition under scale variation, background distraction, and fine-grained inter-class ambiguity. The proposed framework provides an effective solution for sports image classification when labeled samples are limited.