Performance Evaluation of Face Recognition Using ResNet Architecture Under Different Lighting Conditions

Abstract

Face recognition is one of the most widely applied technologies in computer vision, particularly in security and automatic identification systems. This study aims to evaluate the performance of the ResNet50 model in facial recognition under varying lighting conditions. The system was developed using the DeepFace Python library and implemented in real-time through a webcam. The dataset consisted of facial images captured under three lighting conditions: bright, normal, and dim. Each image was processed through face detection, alignment, and normalization stages to fit the model’s input format. The evaluation measured accuracy, precision, recall, F1-score, and average recognition time. Experimental results show that lighting significantly affects recognition performance. The highest accuracy of 98% was achieved under bright lighting, while performance dropped to 84% under dim conditions. The average recognition time ranged between 140–160 milliseconds, indicating that the system performs efficiently in real time. Overall, the ResNet50 architecture demonstrates strong capability in facial recognition under optimal lighting but requires improvement in preprocessing to handle extreme illumination variations.