Deep Learning-Based NIR Face Detection under Adverse Illumination with Explainable AI

Abstract

Conducting face detection in low-light or nighttime conditions is quite difficult for purposes such as surveillance, security, and low-light imaging systems. This paper presents a YOLOv12-based deep learning pipeline that aims at face detection under quite challenging lighting conditions. The illumination variations being a challenge in face recognition, the model was trained on a diverse dataset covering in addition blur and conversion into grayscale, and CLAHE (Contrast Limited Adaptive Histogram Equalization). Additionally, to fine-tune the model’s hyperparameters, a hybrid optimization approach combining Harris Hawks Optimization (HHO) and Whale Optimization Algorithm (WOA) is employed, improving detection accuracy and efficiency. The proposed system achieves 0.994 precision, 0.944 recall, and 0.991 mAP50, demonstrating its high performance even in low-light conditions. In addition to providing a degree of model explainability and interpretability, XAI techniques such as LIME and Feature Map Visualizations aid in interpreting the important areas of influence on the model's decisions for face detection, thus building trust and interpretability. The proposed model is highly suitable for real-time deployment for security and monitoring applications.