Low-light image enhancement by deep learning network for improved illumination map

Abstract

In the case of insufficient illumination conditions, the quality of the captured image is poor. At this time, increasing the brightness of the dark region of the whole image will inevitably aggravate noise pollution. Therefore, the ideal state of low-light image enhancement should be to make the dark areas brighter while not allowing the noise part to be enhanced as well. In this paper, we combine Retinex theory and convolutional neural networks to build a simple and effective model with three sub-modules: the decomposition module, illumination module, and reflection module, in which the decomposition module is used to decompose the input low-light image into illumination and reflection maps, the illumination module improves the feature extraction part of the simple convolutional layer used previously and constitutes a new feature extraction structure by Depthwise separable convolution, which makes up for the shortcoming that some dark areas are not well brightened and better adjusts the brightness of the illumination map, and the reflection module adjusts the texture details more clearly because it adds the illumination map as a reference. In this way, the original space is decoupled into two smaller subspaces to the extent that better results can be achieved. We demonstrate through extensive experiments the effectiveness of the design and its superiority over state-of-the-art techniques, especially in terms of robustness to severe visual defects and flexibility in adjusting luminance.