Design of an intelligent evaluation model for university physical education teaching management based on a feature-fusion attention mechanism
Abstract
This study proposes an intelligent evaluation model for university physical education teaching management based on a feature-fusion attention mechanism, aiming to enhance evaluation accuracy and computational efficiency. A perception-enhancement module is first constructed by integrating a shifted-window attention mechanism with a cross-stage partial feature-aggregation module. The hierarchical design, incorporating layer normalization, window multi-head self-attention, and shifted-window multi-head self-attention, strengthens global feature perception. A context-guided feature-fusion network is then developed to adaptively integrate local and regional features, with channel attention used to optimize fused representations. Finally, a CNN-Transformer cooperative architecture embeds shifted-window attention into convolutional layers, preserving convolutional advantages for local feature extraction while improving global modeling. Experiments on UCF101 and Sports-1M show that the proposed model achieves superior precision, recall, F1-score, and mean Average Precision (mAP) compared with baseline methods. On UCF101, precision and F1-score reach 0.768 and 0.683, respectively, while on Sports-1M they reach 0.711 and 0.628, respectively. The model contains only 12.6M parameters, achieving competitive accuracy while reducing computational complexity.