SID-Net: Satellite Image Dehaze Network Using Vision Transformers

Image dehazing is a low-level vision task which aims to generate latent haze-free images given their corresponding hazy counterparts. Past few years have seen Convolutional Neural Network (CNN) based methods dominate both low-level as well as high-level computer vision tasks. However, Vision Transformers (ViT), with their self-attention mechanism designed for images have recently made a breakthrough in highlevel and low-level vision tasks. CNN and Transformer based networks have been proved to be really efficient in performing image dehazing given that certain adjustments and changes are made to their architectures. In this work, we propose Satellite Image Dehaze Network (SID-Net) using CNN and Vision Transformer which integrates a modified ViT block in our CNN based network to dehaze the satellite images. SateHaze1k dataset is used to conduct all the experiments. Our study includes the comparison of the performance of our proposed model with state-of-the-art CNN based dehazing models. The proposed model achieves Peak Signal-To-Noise Ratio (PSNR) score of 20.5695 and Structural Similarity Index Measurement (SSIM) of 0.8642 on thickly hazed images. In future, we would also like use corresponding SAR (Synthetic Aperture Radar) data as condition to the optical data to improve the performance of our proposed model.