Abstract
Although image restoration methods have achieved great success on single distortion, their performance declines when faced with hybrid distortion effects in real-world scenarios. Recently, some methods for hybrid distortions have been explored. Suppose the total type of distortion is N, the types of distortion that the image suffers may change from 1 to N in real scenarios. However, existing methods fail to simultaneously address variations in distortions. Toward this end, a general network architecture named joint distortion estimation and removal network (JDERNet) is proposed for versatile hybrid-distorted image restoration. The distortion information of the hybrid-distorted image is mathematically represented as a distortion vector, and a neural network is constructed to estimate it. The distortion vector is utilized as a prior to modulate the image reconstruction process. Specifically, a novel feature modulation convolution block (FMCB) is designed. In the FMCB, the distortion vector is employed to modulate the feature maps generated by the attention branch (AB) and the non-AB (NAB), thereby enhancing the restoration of images affected by various types of distortions. Experimental results indicate that the proposed JDERNet achieves state-of-the-art performance in restoring images with varying types of distortions, ranging from multiple hybrid distortions to single distortions.