Development of a deep learning-based model for head-scoring classification in Taekwondo Kyorugi

This study aimed to develop a deep learning–based automatic classification model for detecting valid head-scoring actions in Taekwondo Kyorugi. A total of 20 publicly available Taekwondo match videos were collected from YouTube, and frames containing head-scoring attempts were extracted and enlarged, resulting in 499 images. Among these, 265 were labeled as valid head-scoring actions and 234 as non-scoring actions. Convolutional Neural Network–based ResNet models and Transformer-based Vision Transformer models were evaluated under both original and augmented data conditions, followed by hyperparameter optimization of the best-performing model. Results indicated that the Vision Transformer (ViT) achieved the highest performance under the original data condition, with a test accuracy of 87.0% and an F1-score of 0.893 after hyperparameter optimization, aligning with prior research suggesting that Transformer architectures demonstrate robust representation learning even with limited datasets. Rotation-based data augmentation improved the performance of ResNet but did not provide additional benefits for ViT, highlighting that augmentation effects differ depending on model architecture. Hyperparameter optimization further improved ViT performance by approximately 4%. Overall, this study empirically demonstrates the applicability of Transformer-based models for automated scoring in Taekwondo and underscores the importance of data characteristics, augmentation strategies, and optimization techniques for enhancing model performance.