Dynamic Residual Illuminating Network for Low-Light Image Enhancement

Abstract

A low-light image is an image that is taken under insufficient lighting conditions or captured by an underexposed camera. Low-light images commonly exhibit characteristics such as reduced contrast, diminished brightness, pronounced noise, and color distortion, which make them difficult to see and process. Low-light Image Enhancement is a task that involves improving the qualities of these images and is dominated by deep learning approaches nowadays. However, most deep learning methods utilize standard convolution with fixed kernels in their CNN architectures, hindering model's capability to process and recover more information. Besides, noise remains challenging in recovering low-light images. To mitigate these issues, we propose a model based on the Deep Lightening Network (DLN), we modify the Dynamic Region-Aware Convolution (DRConv) and serve them as the encoder and decoder to process more intrinsic features, we modify Residual in Residual Dense Block (RRDB) to a lighter one to reduce noise and alleviate computational burden. Experiments demonstrate the advantages of our proposed model in quantitative evaluation and visual quality.