Low-FaceNet: Face Recognition-driven Low-light Image Enhancement

Images captured in low-light conditions often induce the performance degradation of cutting-edge face recognition models. The missing and wrong face recognition inevitably makes vision-based systems operate poorly. In this article, we propose Low-FaceNet, a novel face recognition-driven network, to make low-light image enhancement (LLE) interact with high-level recognition for realizing mutual gain under a unified deep learning framework. Unlike existing methods, Low-FaceNet uniquely brightens real-world images by unsupervised contrastive learning and absorbs the wisdom of facial understanding. Low-FaceNet possesses an image enhancement network that is assembled by four key modules: a contrastive learning module, a feature extraction module, a semantic segmentation module, and a face recognition module. These modules enable Low-FaceNet to not only improve the brightness contrast and retain features but also increase the accuracy of recognizing faces in low-light conditions. Furthermore, we establish a new dataset of low-light face images called LaPa-Face. It includes detailed annotations with 11 categories of facial features and identity labels. Extensive experiments demonstrate our superiority against the state-of-the-art methods of both LLE and face recognition even without ground-truth image labels. Our code and dataset are available at https://github.com/fanyihua0309/Low-FaceNet.