Improved vision transformer for sports image classification via transfer learning and knowledge distillation

To address the challenges posed by complex visual characteristics in sports image classification, this study proposes an enhanced Vision Transformer (ViT) model built upon the Native Resolution ViT (NaViT) architecture. The principal innovations include replacing conventional single-layer feedforward networks with parallel feedforward networks and introducing an attention-guided distillation framework with input-dependent attention-weighted feature fusion, which adaptively regulates the contribution of frozen distillation features to the prediction pathway. The proposed approach leverages ImageNet-pretrained weights and performs domain-specific fine-tuning on sports datasets. Experimental results demonstrate substantial performance gains. On the multi-sport classification dataset, the proposed model achieves an accuracy of 0.9752 and a Macro-F1 score of 0.9753, outperforming the evaluated CNN benchmarks and ViT variants. For soccer match event classification, it achieves an accuracy of 0.9863 and a Macro-F1 score of 0.9718, while ranking first in the TOPSIS evaluation. These results demonstrate the effectiveness of integrating parallel feedforward processing, transfer learning, and distillation-guided feature fusion for sports image classification under complex visual conditions.