RETINEX-BASED LOW-LIGHT IMAGE ENHANCEMENT WITH EDGE-AWARE LEARNING

Abstract

Improvement of low-light images is an essential and practically relevant problem in the domain of computervision, havingpractical ramifications on the operation of self-navigating vehicles, smart surveillance cameras, medical imaging, and mobile camera phones. Low light images are characterized by a set of degradation artifacts that includes extremely low contrast, because of inadequate photon collection, excessive noise levels, color distortion due to erroneous white-balance adjustment, and the loss of fine detail structures. The main contributions of this paper are introduced in RLEA-Net, a Retinex-based deep learning framework that utilizes physics-motivated Retinex decomposition along with an explicit edge-aware learning mechanism. The RLEA-Net takes an input low-light im age and decomposes it into its illumination and reflectance layers by a learned network, Decom-Net. In addition, an Illumination Enhancement Network, denoted as Illum-Net, uses multi-scale dilated convolutional layers to predict a residual map on top of the es timated illumination map. On the other hand, the reflectance layer undergoes denoising and fine texture restoration through a Reflectance Restoration Network (Ref-Net) archi tecture based on the U-Net architecture. However, the core innovation lies in designing an Edge-Aware Module (EAM), which utilizes Sobel kernel initialized anisotropic filter banks as a fixed-kernel depthwise convolutional layer to generate channel-wise edge maps that influence the feature transformation within Ref-Net. One of the important principles in the design is that of gradient isolation. In other words, the loss related to reconstruction (such as Charbonnier, VGG perceptual, SSIM, and gra dient fidelity losses) are only computed for the Illum-Net and Ref-Net, but DecomNet is trained through a loss term called Retinex consistency, which is based on synthetic targets created analytically from the reference image. This ensures that there is no decomposi tion collapse problem faced by previous networks. The unique loss is gradient-weighted illumination smoothness loss. Tests performed using the LOL standard dataset (485 training and 15 test examples) show that RLEA-Net obtains PSNR of 20.41dB and SSIM of 0.796 for eval15, having around 8.86 million trainable parameters. Although the results themselves are not impressive in comparison with the biggest published architectures, the performance outdoes RetinexNet by 3.64dB while being close to Zero-DCE and KinD in terms of edge fidelity. An ab lation study proves that all three contributions of RLEA-Net contribute equally to the performance — EAM (+1.24dB), gradient-based illumination loss, and Retinex consis tency supervision.