Contrastive Learning for Unsupervised Single Image Dehazing via Frequency Compensation

Most existing dehazing algorithms establish supervised methods through the utilization of synthetic datasets. However, the domain differences between synthetic and real data make it challenging to apply such algorithms to real foggy conditions. In the present study, an innovative unsupervised single image dehazing network is explored. The network achieves one-sided image transformation through contrastive learning, eliminating the need for paired training images and unnecessary generators. Considering the importance of texture, edges, and other information in subjective perception and other visual tasks, a multi-scale frequency information compensation module is developed. This module combines convolutional neural networks with wavelet transforms to facilitate the extraction of detailed information from foggy images. Additionally, to avoid subjective visual artifacts resulting from the excessive use of pixel differences, the letter proposes extracting pseudo-supervisory information from foggy images to assist in recovering details.