This paper presents a comparative analysis of convolutional neural network (CNN) architectures for detecting punches in boxing videos, to improve sports performance analysis through computer vision. The study evaluates the effectiveness of four CNN models, custom CNN, ResNet50, Inception v3, and VGG16, in identifying punches, considering factors such as classification precision, computational efficiency, and robustness to real-world challenges such as class imbalance. The custom CNN architecture, developed in previous work, was shown to provide a good balance between accuracy and computational demands. In contrast, ResNet50 performed well in complex scenarios, demonstrating strong feature extraction capabilities, while Inception v3 demonstrated superior efficiency in handling varying input sizes. VGG16, although effective, proved computationally expensive for real-time applications. The models were evaluated using metrics such as balanced accuracy and F1 score, addressing the issue of class imbalance where punches are less frequent (approximately 3%) than non-punch frames. The experiments were performed using a publicly available boxing punch classification dataset and source code, both published by the authors to facilitate reproducibility and further research. The results indicate that CNNs offer promising solutions for automated punch detection, with implications for broader applications in sports analytics, including training, injury prevention, and tactical analysis. Future research directions include exploring more advanced CNN architectures, incorporating temporal information, and improving real-time processing capabilities. This work contributes to the development of efficient, scalable and reliable computer vision systems for the evaluation of sports performance.