Recurrent Exposure Generation for Low-Light Face Detection

Face detection from low-light images is challenging due to limited photons and inevitable noise, which, make the task even harder, are often spatially unevenly distributed. A natural solution borrow idea multi-exposure, which captures multiple shots obtain well-exposed under conditions. High-quality implementation/approximation of multi-exposure a single image however nontrivial. Fortunately, as shown in this paper, neither such high-quality necessary since our face detection rather than image enhancement. Specifically, we propose novel Recurrent Exposure Generation (REG) module couple it seamlessly with Multi-Exposure Detection (MED) module, thus significantly improve face performance by effectively inhibiting non-uniform illumination noise issues. REG produces progressively efficiently intermediate corresponding various exposure settings, pseudo-exposures then fused MED detect faces across different lighting The proposed method, named REGDet, first ‘detection-with-enhancement’ framework for detection. It not only encourages rich interaction feature fusion levels, but also enables effective end-to-end learning component be better tailored Moreover, clearly experiments, can flexibly coupled detectors without extra low/normal-light pairs training. We tested REGDet on DARK FACE benchmark thorough ablation study, where outperforms previous state-of-the-arts significant margin, negligible parameters.