Logarithmic Retinex Decomposition-Aided Convolutional Neural Networks for Low-Light Image Enhancement

Abstract

Maritime accidents kill thousands of lives every year because of the difficulty in rescuing. Usually, the weather conditions at sea are very complicated, especially at midnight. The images captured by the monitor often suffer from low visibility under this circumstance, which gives rescuers a great challenge to seek survivors. Existing image enhancement methods are not suitable for the offshore environment. We propose an image enhancement model to restore the image from the low-light condition to solve this problem. We found that if we transform an image into a logarithmic image, it can provide more details in darkness, and we decide to make better use of this information because the survivors’ bodies may be hidden in darkness. We build networks based on previous Retinex-based network researches and use the logarithmic image as the original dataset. The De-Net decomposes the image into the reflectance map and illumination map for further adjustment. The recovered logarithmic image will turn back to the output image after the I-Net and R-Net. Experiments have been implemented on synthetic and realistic low-light images to verify the effectiveness of our method. Experimental results have illustrated that it is widely applicable to maritime low-light images.