Multiscale Synergism Ensemble Progressive and Contrastive Investigation for Image Restoration

Image restoration refers to enhance the visibility of degraded images. Given the complex and variable aspects of image degradation, current methods tend to sacrifice contextual information in their effort to remove as much degradation as possible. To strike a balance between detail preservation and degradation recovery, we develop a synergism ensemble progressive and contrastive (SEPC) network to address image restoration, which explores the complementarity of single-scale and multiscale learning to achieve the restoration with semantic awareness and detail preservation. Specifically, deep contextual information derived from the multiscale branch aids in the estimation of an attentive mask. The interference is pinpointed in this mask and subsequently directed to the single-scale branch for detail-preserving restoration. To lessen the dependency on deep features across different levels, we unfold a communal unit progressively to accommodate the harsh interference. Acknowledging the discrepancy between synthetic and real-world scenarios, we incorporate contrastive learning to push the results toward interference-free images, and distance them from their degraded counterparts simultaneously, which enhances the generalization across a range of restoration tasks. To evaluate the effectiveness of our method, we apply it to single image deraining, denoising, and demosaicing. Extensive experiments deliver superiority against state-of-the-art methods quantitatively and qualitatively.