TransDehaze: transformer-enhanced texture attention for end-to-end single image dehaze

Image dehazing plays a pivotal role in enhancing the performance of computer vision applications by restoring the original colors and textures of images affected by haze. Traditional dehazing methods, primarily based on atmospheric scattering models, often lead to image distortions due to imprecise parameter estimations. The emergence of deep learning has shifted the paradigm toward neural network-based approaches for image dehazing. However, these approaches, largely relying on localized convolution operations, frequently result in the loss or distortion of texture features in the dehazed images. In response to these limitations, this paper presents TransDehaze, an innovative dehazing algorithm that effectively integrates the strengths of Transformer and U-net architectures to better preserve image textures. TransDehaze leverages multiple transformer structures to extract features at various depths, subsequently fusing these features using learnable weights to accurately restore texture details. Additionally, the algorithm is enhanced with power normalization and cross-link techniques to optimize transformer’s efficacy. Our comprehensive evaluation involves a range of mainstream dehazing methods and three distinct datasets: NYU-Depth V2, IST, and TOST. The results demonstrate that TransDehaze not only surpasses existing state-of-the-art dehazing algorithms in terms of dehazing quality but also significantly improves detection efficiency. An ablation study is included to highlight the individual contributions of each module within TransDehaze, illustrating its remarkable advancements in image dehazing. Our code is at https://github.com/igoindown/TransDehaze.git .