ZRRD-MBNet：Zero-Reference RETINEX Decomposition-Based Multi-Branch Network for Low-Light Image Enhancement

Abstract

Abstract To address the challenges of low visibility, poor contrast, and significant noise in images captured under various extreme conditions, such as backlighting and low-light situations, we propose a novel zero-reference Retinex decomposition-based multi-branch network known as ZRRD-MBNet for enhancing low-light images. The ZRRD-MBNet is divided into two main components: decomposition and recovery. The decomposition part can separate the input low-light image into three distinct components: the illuminance map, the reflection map, and the noise map. On the other hand, the recovery part follows a dual-path approach. The first path involves gamma transforming the generated illuminance map to enhance brightness effectively. Simultaneously, the second path divides the input low-light image element-wise by the originally decomposed illuminance map. The result is then subtracted element-wise from the noise map obtained during decomposition, yielding a denoised reflection map. To obtain the final enhanced image, the denoised reflectance map is multiplied with the brightness-enhanced illuminance map. The specific loss functions are critical for updating the weight parameters of ZRRD-MBNet and guiding the network during low-light image decomposition. Extensive qualitative and quantitative experiments are conducted using publicly available datasets. The results demonstrate that our method produces high-quality images with clarity, minimal noise, and absence of artifacts.