A DEEP LEARNING FRAMEWORK FOR REAL-TIME STRESS DETECTION USING FACIAL EXPRESSIONS

Abstract

Given its direct effects on human health productivity at work and general well-being stress detection has emerged as a crucial area of affective computing research. Self-reports and physiological sensors which are either subjective invasive or expensive to implement at scale are frequently used in traditional stress assessment methods. This study suggests an Enhanced VGG-Based Approach for Facial Expression-Driven Stress Detection (EVSD-Net) which uses subtle facial cues to accurately identify stress levels in order to overcome these limitations. By adding lightweight convolutional blocks to eliminate unnecessary parameters a hybrid channel–spatial attention mechanism to highlight stress-relevant areas (like wrinkles on the forehead tight lips and strained eyes) and a fine-tuned transfer learning strategy initialized with pretrained ImageNet weights the suggested framework enhances the standard VGG16 architecture. Both binary (stressed and non-stressed) and Softmax classifiers are supported by a hybrid classification layer that integrates fully connected (FC) layers. non-stressed) and multi-class stress classification (high medium and low). Preprocessing and augmentation of facial datasets improved feature extraction using the modified VGG network dimensionality reduction and regularization for better generalization and Softmax classification comprise the four main stages of the methodology. Accuracy Precision Recall F1-score and AUC metrics were used to evaluate the experimental results using four benchmark facial expression and stress-related datasets. The results show that EVSD-Net outperforms baseline models like standard VGG16 ResNet18 and MobileNetV3 with an overall accuracy of 97.84%. The superiority of the suggested model is further confirmed by a comparison with previous research which shows significant improvements in detection accuracy and robustness under various illumination pose and demographic circumstances