A Multi-Layer Perceptron with Logistic Regression Based on Residual Networks for Martial Arts Action Recognition

Nowadays, Martial arts embody a dynamic fusion of physical discipline, mental toughness, personal development, community, and cultural heritage to cultivate a holistic approach to well-being and modern living. Precise identification of martial arts techniques is essential to ensure safety, optimize performance, and foster personal growth, whereas inaccurate recognition can spread harm, aggression, and cultural misconceptions. Various methods have been proposed previously to recognize the actions of martial arts. To overcome these problems, a Multi-Layer Perceptron with Logistic Regression (MLP-LR) is proposed to recognize the actions of martial arts. At inception, an Ultimate Fighting Championship/Mixed Martial Arts Fights Image Segmentation (U/M-FIS) dataset is considered as an input for recognizing martial arts actions. For preprocessing, image resizing, image normalization, data augmentation, noise reduction, and contrast enhancement are employed. Residual Network 50 (ResN et50) is considered for feature extraction to reduce feature dimensionality, and to enhance robustness to noise and variability. Next, data preparation is done with training, testing, and pattern validation which is followed by classification using MLP-LR. The proposed MLP-LR model attained higher performance metrics such as accuracy, precision, recall, and F1 score of 99.96%, 99.87%, 99.52%, and 99.47%, when compared to other existing methods like Convolutional Neural Network (CNN).