Cricket analysis is crucial for coaches, analysts, and players to get various insights to improve training, performance, and viewer engagement. The present study identifies and classifies seven cricket shots, such as cover, pull, and sweep, from video input. Different approaches have been used to obtain key frames from the input and send them to several deep-learning architectures. A base model is built using CNN-GRU and compared with other architectures like MobileNet and ResNet-3D. A higher accuracy of 93.94% is obtained through ResNet-3D. Integration of Physics Loss gave an additional 3.8% improvement in model performance and 20% reduction in training time through Physics-Informed Neural Networks. The ResNet-3D model with physics-informed loss (ResNet-PINN) achieved the highest accuracy of 97.73%, a 5.56% improvement over the base model, and a 68.48% reduction in training time. This efficiency gain is attributed to the physics-informed constraints, which guide the model toward physically plausible solutions and help learn cricket dynamics efficiently. These findings demonstrate that integrating domain-specific physics into neural networks not only improves classification accuracy but also enhances computational efficiency, making it viable for real-time applications.