Automated Serve Shot Classification in Table Tennis Using CNN-LSTM Framework

Analysis of table tennis serve shot is significant for player strategies and performance. In this work we propose a hybrid deep learning framework based on Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks that aims to automate serve classification and success prediction. A custom dataset of 10,000 annotated serve shots was used for training and evaluation of the proposed model, which obtained a serve class accuracy of 92.3%, which was much superior to baseline CNN (85.6%) and LSTM (80.2%) models. In addition, the recorded success prediction accuracy was 88.7%, which proved the feasibility that the model can realize to learn to predict effective serves with high precision. For evaluation metrics, the model's precision was 0.91, recall was 0.93, and f1 score was 0.92, validating it to separate among different serve types. The overall performance of the framework in predicting serve success and failure can be confirmed by the area under the ROC curve (AUC) at 0.93. Results from confusion matrix analysis confirmed that the model could classify frequent serve types with greater than 95% accuracy but slight misclassifications occurred with visually similar serve types. As Training and Validation accuracies improve steadily to 92.3% and 91.5% respectively and losses fall from 0.6 to 0.12, measure of learning, the lack of fit is minimal. The results show that the CNN-LSTM framework can be used to provide information to players and coaches, empowering data driven decision making and performance optimization. As future work, the dataset will be extended, the model architecture will be further refined, and realtime analytics will be added for improved practical usefulness in table tennis coaching and strategic development.