ResNet50-Enhanced Multi-Layer Perceptron with Logistic Regression for Precise Martial Arts Action Recognition

Martial arts integrate physical training, mental resilience, personal growth, social bonding and cultural heritage, offering a holistic wellness solution for modern life. Accurate recognition of martial arts techniques is critical to ensure safety, optimize athletic performance and promote cultural understanding, while misrecognition may cause avoidable injuries, aggressive conduct and biased cultural perceptions. Despite existing action recognition methods for martial arts, challenges like weak environmental robustness and limited generalizability remain unsolved. To address these gaps, this study proposes a hybrid MLP-LR model enhanced by ResNet50 for feature extraction, using the U/M-FIS dataset (5,000+ annotated images, 10+ classes) as input. A rigorous preprocessing pipeline (resizing, normalization, augmentation, denoising, contrast enhancement) is adopted to improve data quality. ResNet50 extracts discriminative features to reduce dimensionality and enhance anti-noise ability, and the MLP-LR classifier performs action classification after dataset splitting. Experimental results show the proposed model achieves superior performance: 99.96% accuracy, 99.87% precision, 99.52% recall and 99.47% F1-score, outperforming traditional CNN methods. This work provides a reliable and efficient framework for martial arts action recognition, with promising applications in sports training, athlete performance analysis and cultural heritage preservation.