Unsupervised Hybrid-distorted Image Restoration Method Based on Feature Disentangled Representation Learning

For the hybrid-distorted image restoration in real scenes, an image restoration algorithm based on unsupervised dual learning is proposed with the Generative Adversarial Networks and the encoder. The algorithm introduces feature decoupling module where the different feature representations from different degradation mechanisms are assigned to different feature channels by modifying the normalization based on gain control so the feature disentanglement in channels is realized by independent feature representations. Meanwhile, the channel attention mechanism is used to realize the restoration of image content feature of clean images. Compared with the SRN algorithm, the proposed algorithm improves 0.499 dB and 0.044 in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) on GoPro dataset. Compared with the OWAN algorithm, the proposed algorithm improves 0.163 dB and 0.015 in terms of PSNR and SSIM on DIV2K dataset. The experiments also demonstrate that the detailed information can be restored by the proposed algorithm.