<![CDATA[Degradation of illumination intensity in low light environments is a major issue that reduces the accuracy of deep learning based face detection and recognition systems. This study aims to optimize the performance of the biometric processing pipeline under extreme low light conditions without retraining the model. The novelty of this research lies in the design of a retrainless integration between the hybrid preprocessing of CLAHE and Bilateral Filter on the luminance channel of the LAB color space with the integrated processing pipeline of YOLOv5-Face and FaceNet. Mass testing was conducted using pair-based evaluation on 3,000 face pairs from the VGGFace2 benchmark dataset, simulated in a controlled manner using a gamma exponent value (y = 0,5). Experimental results show that the preprocessing stage successfully restored the YOLOv5-Face Detection Rate from 88.7% to 89.8%. Meanwhile, in the identity verification stage, the FaceNet model recorded an increase in class separability, achieving the highest Area Under the Curve (AUC-ROC) value of 0.927 (classified as excellent), a global accuracy of 89.3%, and the ability to maintain the stability of the Cosine Distance Gap at an index of 0.6005. This characteristic proves the robustness of the feature vector geometry in separating boundaries between identities without overlapping. The implementation of the system into a Streamlit web application confirms that this traditional contrast restoration method remains relevant, efficient, and reliable for securing biometric verification under low-light conditions.]]>
