Pose-guided cross-scale attentional transformer learning with spatial–contextual fusion for explainable frame-based tennis stroke recognition

Understanding sports actions from visual data has become increasingly important for intelligent coaching systems, athlete performance analysis, and automated video analytics. Tennis presents a particularly challenging problem because many stroke categories differ only in subtle body poses, racket positions, and spatial interactions. Existing convolutional and transformer models often struggle to capture both fine-grained local patterns and broader contextual information at the same time. To address these limitations, this study proposes a novel pose-guided cross-scale attentional transformer framework with spatial–contextual fusion for explainable tennis stroke recognition. The proposed model introduces a Pose-Aware Feed-Forward Network (P-FFN), a Cross-Scale Gating Mechanism, and a Dual-Branch Spatial–Contextual Fusion module to improve the learning of discriminative action features. Multi-scale convolutional embeddings are combined with transformer-based global attention, allowing the network to jointly model pose structures and contextual semantics related to tennis stroke dynamics. Experiments conducted on a balanced dataset of four tennis stroke classes demonstrate that the proposed approach achieves an accuracy of 96.28%, outperforming conventional CNN and transformer baselines. Seven-fold cross-validation and 95% confidence intervals further confirm the robustness and consistency of the model. To enhance transparency, Grad-CAM, LIME, and SHAP are employed to visualize and quantify feature importance. The model attains a Grad-CAM focus ratio of 91.4% and a SHAP relevance score of 93.6%, indicating that it consistently attends to meaningful regions associated with player actions. The findings show that integrating pose guidance, cross-scale learning, and explainable AI provides an effective and trustworthy framework for fine-grained tennis stroke recognition. Not applicable.