Low-light Image Enhancement with Deep Blind Denoising

Abstract

Visual processing technology based on the high-quality image has been widely applied to maritime supervision. However, maritime accidents easily occur in a low-light environment. The images captured by the monitor often suffer from low visibility and noise, which will directly affect the rescue efficiency and accident investigation. To improve the quality of low-light images, we propose a low-light enhancement method based on the Retinex theory and deep blind denoising. In particular, the coarse illumination map is estimated via finding the maximum in the three channels of the original images. Then, we adopt the Guided filter to refine the coarse illumination map and adjust it by Gamma correction. Based on the Retinex theory, the reflectance map can be easily obtained by the original image and the refined illumination map. Meanwhile, the convolutional neural network is introduced to blindly remove the noise of the reflectance map. Finally, we can obtain the enhanced image by the adjusted illumination and denoised reflectance maps. Experiments have been implemented on realistic and synthetic low-light images to verify the effectiveness of our method. Experimental results have illustrated that it is superior to several popular low-light enhancement methods.