Multi-stage dehazing network: Where haze perception unit meets global and local progressive contrastive regularization

Image dehazing is a crucial low-level restoration task that aims to recover a clear image from a hazy observation. Recent learning-based approaches have demonstrated impressive performance in this area. However, there are still two drawbacks: (1) Existing dehazing architectures do not sufficiently consider non-uniform haze distribution, indicating that the haze location information is underexplored. (2) Naïve contrastive regularization fails to provide enough constraint force in solution space, i.e., negative-oriented supervision information cannot be fully utilized during the training stage. Consequently, we establish a Multi-stage Dehazing Network (MSD-Net) to achieve single image haze removal. For one thing, we build a haze perception unit (HPU) based on a self-calibration attentive paradigm. This unit can effectively encode haze location information as prior guidance and further enhance its feature representation capabilities. For another, we tailor a global and local progressive contrastive regularization (GLPCR) to explore negative-oriented supervision information. Specifically, the negative samples are derived not only from the original hazy images but are also progressively updated through the pseudo-restoration results of the multi-stage architecture. To tackle the learning ambiguity arising from diverse negative samples, we employ a curriculum learning strategy during the training phase. Moreover, our GLPCR operates in both global and local manners, encouraging the network to retain rich information from both image-wise and patch-wise perspectives. Extensive experiments demonstrate that our MSD-Net can achieve remarkable dehazing performance compared with other state-of-the-art methods on several commonly used hazy dataset benchmarks.