Noise-aware Zero-Reference Low-light Image Enhancement for Object Detection

Computer vision advancement has proven to be able to automate many practical tasks such as object detection and recognition in challenging environments. However, most notable computer vision models are optimized to work under environment with ideal lighting conditions. Real-world scenarios are uncontrolled and it is common to encounter quality and performance deterioration due to challenges in poor lighting, especially related to the amplification of noise signals. Consequently, there is an increase of vision enhancement related works to maintain such models' performances, however there is still a gap in exploring the practical implication that such existing enhancement work has on detection models, as well as the issue and handling of noise signals. Hence, this paper investigates the incorporation of noise information into the enhancement modelling with the specific aim to improve the performance of object recognition. Building upon the zero-reference deep curve estimation (Zero-DCE) approach, a noisy data training strategy is designed in order to introduce noise priors into the curve estimation training to produce better image structure enhancement. Furthermore, the commonly implemented post-process denoising approach is also studied in this work to find out its impact and effectiveness in the context of object detection. Experiments on the ExDark dataset show that the enhanced images produced by the proposed approach is able to improve object detection in low-light images using YOLOv5 with an increase of up to 3% in precision, and comparable performance with significantly more complex state-of-the-art low-light image enhancement models.