MDenseNet121-ISOA: An Innovative Framework for Revolutionizing Human Action Recognition in Thermal Images for Enhanced Accuracy and Real-Time Performance

The ability to identify emotions and facial expressions is necessary for applications like security monitoring healthcare tracking and human–computer interaction. However current methods frequently fail to maintain consistent recognition accuracy in situations with noise lighting changes obstacles and subtle facial or movement patterns especially in real-time thermal imaging. This study offers a thorough framework that combines optimized techniques for preprocessing feature extraction localization classification and hyperparameter adjustment to recognize human emotions and actions from thermal images in order to solve these issues. The suggested preprocessing method uses min-max normalization and contrast enhancement to draw attention to subtle facial expressions and movement cues after applying median filtering to reduce noise while maintaining emotion-related details. An improved Res2Net architecture is used to extract multi-scale hierarchical features enabling efficient representation of intricate spatial and contextual patterns. To achieve accurate region detection in complex environments an improved YOLOv9 model that incorporates enhanced bounding-box regression and anchor-free detection is used for emotion and action localization. To improve recognition performance a modified DenseNet-121 classifier with enhanced gradient propagation and dense feature reuse is applied. Additionally an Improved Snake Optimization Algorithm (ISOA) which increases convergence efficiency and guarantees stable performance under various environmental conditions is used to optimize model hyperparameters. The suggested MDenseNet121-ISOA framework greatly outperforms current methods according to experimental evaluation on several thermal image datasets. In particular the model attains 99.518% accuracy, 99.51% precision, 99.56% recall, and 99.383% F1-score on the IITR-IAR dataset 99.565% accuracy, 99.53% precision, 99.41% recall, and 99.594% F1-score on the LTIR dataset and 99.51% accuracy, 99.96% precision, 99.564% recall, and 99.762% F1-score on the InFAR dataset. These findings show that the suggested framework is appropriate for real-time human emotion and action recognition in difficult thermal imaging environments since it offers better recognition accuracy at a lower computational cost than current approaches.