Contrastive adaptive frequency decomposition network guided by haze discrimination for real-world image dehazing

Recent unsupervised image dehazing methods used unpaired real-world training data for enhancing generalization on real-world scenes. However, these methods often require dehazing and rehazing cycles with auxiliary networks for training, resulting in high computational costs and extended training time. In this work, we propose an unsupervised dehazing framework called Contrastive Adaptive Frequency Decomposition Dehazing Network (CAFDD). By incorporating carefully designed network structure and constraints, our CAFDD well avoids additional training overhead and needs only 1.91M parameters. Specifically, we first consider the following insights, including: 1) Haze primarily affects high-frequency components in an image, resulting in blurred edges; 2) Low-frequency components capture the large-scale variations with less susceptibility to haze; and 3) Existing unlearnable frequency decomposition methods such Fourier transform often suffer from information loss, and thus develop the novel PMP (Pointwise convolution-Max pooling-Pointwise convolution) and DAD (Depthwise convolution-Average pooling-Depthwise convolution) blocks to automatically extract high and low-frequency features from input images for accurately estimating transmission map. Then, we propose haze discrimination (HD), a new pretext task for contrastive learning in image dehazing, by forming positive and negative pairs based on haze presence, in order for guiding the network to extract visibility-related features. Last, to get rid of the rehazing cycle and improve training efficiency, we construct a pixel-level constraint, histogram equalization-based texture loss function, which enhances the sharpness and realism of the generated images. Through extensive experiments, we demonstrate the superiority of our CAFDD over the state-of-the-art dehazing approaches on real-world land and overwater images.