Pixel Transformer for Synthetic-to-Real Single Image Dehazing
Yuting Zhang, Fan Wang, Dong Yin · 2023
Existing dehazing methods are always unsatisfactory on real images. In this paper, we propose a novel dehazing model Pixel Transformer with an encoder-decoder structure based on the self-attention mechanism. First, a non-homogeneous synthetic haze dataset, NHS-Haze, is constructed using a random sampling stitching method to simulate real images. Second, to capture the similarity of the internal structure of the image, an encoder consisting of two basic blocks accompanied by a patch merging layer is designed. Also, a pixel shuffle layer is added after each basic block in the decoder to reconstruct the details from the feature maps. Third, the NHS-Haze dataset is used to decrease the domain difference between the real and synthetic images to mitigate the domain drift problem. Finally, cross-testing is performed on the public, synthetic dataset Haze4K and SOTA methods to evaluate the robustness of each dehazing model on different datasets under realistic scenarios.