Image Restoration Methods: A Comprehensive Overview of Single and Multiple Degradation Approaches

According to the focus of this paper, the existing image restoration methods could be classified into two families, i.e., image restoration for single (IRSD) and multiple degradations (IRMD). \nImage Restoration for Single Degradation: IRSD aims to recover a clean image from the degraded observation which is corrupted by only a specific degradation type with a fixed corruption ratio. For instance, as one of pioneering deep denoising methods, DnCNN [53] cannot handle the multi-degradation case even be failed when the noise ratio is unseen during training. Other image restoration tasks have also faced the similar challenge, such as de-bluring [2,12,29,32\u201334,36], deraining [10,17,24,42,46,49,50,52], and dehazing [1,15,20,25,28,35,37,38]. In recent, some works [13,26,39,51] show certain generalizability to different degradations. However, they need to train different models for different degradations, which are not all-in-one solutions as expected in practice. \nImage Restoration for Multiple Degradations: Recently, there are some works [3, 23] shift their attention to IRMD by adopting a multi-input and -output network structure. For example, Li et al. [23] proposed an all-in-one model to handle multiple bad weather degradations (e.g. rain, fog and snow) and each degradation is specifically tackled by an encoder. Chen et al. [3] proposed a transformer-based image restoration method which handles multiple-degradations by using an architecture of multi-heads and multi-tails. The most similar method with our approach may be [8]. However, the method still needs to know some priors of the input (e.g., noise ratio and JPEG quality) for parametrizing the network in a meta-learning manner. To summarize, although the above methods have stepped towards IRMD, they still require the degradation information in advance so that the input could be sent into the corrected head or the meta information could be generated.