Enhanced multibranch Taylor transformer for real-world image de-fogging method

In recent years, the Transformer network has shown great potential in the field of computer vision by virtue of its global sensing field and flexible adaptability to inputs, and widely applied in defogging. However, there are problems of loss of details and lack of contrast in low-light images after defogging. To solve this problem, based on Multi-branch Taylor Former, firstly, the image enhancement algorithms Multi-scale Retinex, Enhanced Atmospheric Scattering Model, and Adaptive Gamma Correction are combined, which improves the overall visual effect of the image; secondly, the non-local module is introduced in the decoding stage, which enhances the integration between global features fusion with local features, thus enhancing the global information capturing ability; finally, the multi-scale feature extraction ability is imposed by optimising the multi-head attention mechanism. The experimental results show that the improved Multi-branch Taylor Former achieves performance improvement: PSNR is increased by 13%, SSIM by 11%, and Entropy by 6.3%.