Retinex low-light image enhancement network based on attention mechanism

Images taken at night, under cloudy conditions, etc., are degraded and of poor quality, which can cause problems such as low brightness, excessive noise, and loss of information. In recent years, deep learning-based methods have led to a significant breakthrough in the challenging task of processing the enhancement of low-light images. We propose a new attention-based network incorporating Retinex theory for the enhancement of low-light images. First, the method in this paper requires pairs of low-/normal-light images for training, and the respective illumination and reflectance maps are decomposed by the first part of the designed network according to the commonality of both. Secondly, an attention mechanism module is inserted in the convolutional layer in the second part of the network, to adaptively adjust the luminance information of the illumination and to preserve the consistency of the image structure. Finally, the new illumination map estimated in the second part is combined with the previous reflectance map to obtain the final enhanced image. The experimental results show that the method achieves better results both quantitatively and qualitatively, with obvious luminance enhancement, less noise, better color reproduction, clearer texture information, and overall superiority compared to existing advanced methods.