Single Image Dehazing Using Quantization Aware Swin Transformer Based Unet Architecture

This Haze arises due to extreme bad weather conditions, and it causes light refraction due to the presence of suspended particles in the air and that reduces the visibility of the distant objects. Light scattering from suspended particles severely limits the visibility range of the objects. Such scenarios rarely happen due to inclement weather conditions and to address this issue, a simple yet effective image dehazing techniques are highly appreciated for next level applications such as lane detection, object detection, scene localization etc. Estimation of depth and airlight are two critical parameters in reconstructing the haze-free image. Many researchers have contributed to estimating the transmission map and airlight separately, and a haze-free image was reconstructed using an atmospheric scattering model. Recent advances in deep learning for modelling highly nonlinear systems have been extremely successful. In this paper, we presented a dehazing model using swin transformer based UNet architecture with long skip connections to improve learning.The performance of the proposed swin transforemer based UNet architecture for image dehazing is compared with the state-of-the-art methods on middlebury and IHAZE datasets in terms of computational complexity, SSIM, PSNR and SSEQ.