Combining attention mechanism and Retinex model to enhance low-light images. (May 2022)

• Record Type: Journal Article
• Authors: Wang, Yong; Chen, Jin; Han, Yujuan; Miao, Duoqian
• Is Part Of: Computers & graphics. Volume 104(2022)
• Page Start: 95
• Page End: 105
• Publication Date: 2022-05
• Subjects: Low-light image enhancement -- Retinex model -- Attention map -- Image decomposition -- Image restoration; Computer graphics -- Periodicals; 006.6
• DOI: 10.1016/j.cag.2022.04.002
• Languages: English
• ISSNs: 0097-8493

Abstract

Abstract: Low-light image enhancement is challenging due to intractable problems such as color distortion and noise, which hide in the dark. Simply enhancing the brightness of dark areas will inevitably amplify hidden artifacts. We have observed more noise in the underexposed areas of images than in the normally exposed areas. Attention mechanism can be used to emphasize the vital information of the processed object and suppress some irrelevant information. Inspired by these observations, we propose a deep network that Combines Attention mechanism and Retinex (CA&R Net) model to enhance low-light images. Firstly, we develop an attention map to evaluate the degree of image underexposure and guide enhancement in a region-adaptive manner. This way, it can enhance underexposed areas and avoid over-enhancing normally exposed areas. Secondly, we use the reconstructed reflectance and low illumination to predict the illumination layers of the image jointly. This joint prediction utilizes the attention mechanism, making illumination adjustment achieve better results. The quantitative experimental results show that the CA&R Net can successfully handle noise, color distortion, and multiple types of degradation with the power of attention information. Moreover, both SSIM and PSNR are better than other advanced methods. Graphical abstract: Highlights: The proposed CA&R Net combines attention and Retinex to enhance low-light images. Develop an attention map to guide the enhancement in a region-adaptive manner. Use restored reflectance and low illumination for jointly predicting to illumination layer. The CA&R Net can handle noise, color distortion, and multiple types of degradations. Extensive experiments have been conducted to demonstrate the superiority of CA&A Net.