A Progressive Fusion Framework of Rotationally-Enhanced 3-D CNN and GCN-DTW for Table Tennis Action Recognition and Quality Evaluation

Most existing studies on table tennis training focus on either action recognition or scoring alone, lacking a systematic way to model both tasks together, which limits the ability to provide complete and reliable feedback for training. To address this issue, a table tennis training action analysis framework is proposed, which integrates rotation-enhanced three-dimensional convolutional neural networks (3D CNN), multi-head spatiotemporal attention (MSTA), and graph convolutional networks combined with dynamic time warping (GCN+DTW). The modules were arranged sequentially for feature extraction, temporal dependency modeling, alignment correction, and structural evaluation, forming a progressive fusion mechanism. The results show that the proposed model achieved over 92% Rank-1 accuracy in action recognition tasks, while the mean absolute error (MAE) in quality scoring is reduced by more than 30% compared with other methods. In the consistency analysis, the weighted DTW improved the agreement with human scores by about 15%–20% on eight typical actions. These results show that the model improves recognition accuracy and scoring consistency, confirming the effectiveness of its joint design. The study provides strong support for the development of action recognition and action quality assessment methods.