Derivative feature and residual spatial attention for low-light image enhancement

Images taken in low-light conditions often have the problem of poor visibility. Besides inadequate lightings, different types of image quality degradation, such as a large amount of noise and color loss due to the limited quality of cameras and camera ISO setting, cause low quality of the captured image. However, directly amplifying the darkness of the lowlight image will inescapably bring into the pollution of the image. Therefore, the task of low-light image enhancement needs to kindle the dark regions and remove image degradation. To achieve this task, our work builds a Retinex theorybased neural network, which decomposes the input images into an illumination map and a reflectance map. Illumination map, representing the light information, is used for brightness adjustment, while reflectance map, representing the color information, is responsible for reconstructing low-light image into enhanced image with adjusted illumination map. However, there are few studies that notice the derivative of the image is used to solve the noise problem in Retinex decomposition and use spatial attention-based residual structures to increase the effect of light enhancement. For Decomposition sub-Network (Decom-Net), we purpose derivative features to alleviate the occurrence of noise in the reflectance map in the process of low-light image decomposition. For Illumination Enhancement sub-Network (Relight- Net), we use the Gaussian blur for reducing the problem of brightness enhancement degradation and build the Residual Spatial Attention Block (RSAB) to enlarge the volume and increase the capability of pixel-to-pixel mapping. Experiments are implemented to shows the effectiveness of our network, which improves the performance of previous methods on a large scale.