Multistage supervised contrastive learning for hybrid-degraded image restoration

Natural image degradation is frequently unavoidable for various reasons, including noise, blur, compression artifacts, haze, and raindrops. The majority of previous works have advanced significantly. They, however, consider only one type of degradation and overlook hybrid degradation factors, which are fairly common in natural images. To tackle this challenge, we propose a multistage network architecture. It is capable of gradually learning and restoring the hybrid degradation model of the image. The model comprises three stages, with each pair of adjacent stages combining to exchange information between the early and late stages. Meanwhile, we employ a double-pooling channel attention block that combines maximum and average pooling. It is capable of inferring more intricate channel attention and enhancing the network’s representation capability. Then, during the model training step, we introduce contrastive learning. Our method outperforms comparable methods in terms of qualitative scores and visual effects and restores more detailed textures to improve image quality.