Learning an Enhancement Convolutional Neural Network for Multi-degraded Images

Abstract

Although image enhancement methods have been widely applied in various outdoor vision systems, the existing methods still face two critical problems. On the one hand, the existing methods only consider a single degradation. However, in practical applications, image quality is usually degraded by multiple factors. The methods designed for the single degradation factor cannot achieve good performance when facing multi-degraded images. On the other hand, the imaging model-based enhancement methods which use prior knowledge or handcrafted features to perform image enhancement may bring some fitting errors. Therefore, considering multiple degradations in images, an image enhancement method is proposed in this paper. Firstly, a new image degradation model based on the multiple scattering model is proposed, which is used to characterize multiple degradations caused by haze, mixed with blur and noise. Then, an image enhancement convolutional neural network (CNN) based on ResNet is proposed to learn the implicit mapping model between low-quality and high-quality images in the pixel domain directly. The CNN network has been trained with an end-to-end learning manner. Experimental results on the synthetic dataset and real-world hazy images verify the superiority of the proposed method, while compared with the state-of-the-art methods.