HAIR: Hypernetworks-based All-in-One Image Restoration

Image restoration aims to recover a high-quality clean image from itsdegraded version. Recent progress in image restoration has demonstrated theeffectiveness of All-in-One image restoration models in addressing variousunknown degradations simultaneously. However, these existing methods typicallyutilize the same parameters to tackle images with different types ofdegradation, forcing the model to balance the performance between differenttasks and limiting its performance on each task. To alleviate this issue, wepropose HAIR, a Hypernetworks-based All-in-One Image Restoration plug-and-playmethod that generates parameters based on the input image and thus makes themodel to adapt to specific degradation dynamically. Specifically, HAIR consistsof two main components, i.e., Classifier and Hyper Selecting Net (HSN). TheClassifier is a simple image classification network used to generate a GlobalInformation Vector (GIV) that contains the degradation information of the inputimage, and the HSN is a simple fully-connected neural network that receives theGIV and outputs parameters for the corresponding modules. Extensive experimentsdemonstrate that HAIR can significantly improve the performance of existingimage restoration models in a plug-and-play manner, both in single-task andAll-in-One settings. Notably, our proposed model Res-HAIR, which integratesHAIR into the well-known Restormer, can obtain superior or comparableperformance compared with current state-of-the-art methods. Moreover, wetheoretically demonstrate that to achieve a given small enough error, ourproposed HAIR requires fewer parameters in contrast to mainstreamembedding-based All-in-One methods. The code is available athttps://github.com/toummHus/HAIR.