DC-GAN with Feature Attention for Single Image Dehazing

Abstract In recent years, the frequent occurrence of smog weather has affected people's health and has also had a major impact on computer vision application systems. Images captured in hazy environments suffer from quality degradation and other issues such as color distortion, low contrast, and lack of detail. This study proposes an end-to-end, adversarial neural network-based dehazing technique called DC-GAN that combines Dense and Residual blocks efficiently for improved dehazing performance. In addition, it also consists of channel attention and pixel attention, which can offer more versatility when dealing with different forms of data. The Wasserstein Generative Adversarial Network with Gradient Penality(WGAN-GP) was used as an enhancement method to correct the shortcomings in the original GAN's cost function and create an improvised loss. On the basis of the experiment results, the algorithm used in this paper is able to generate sharp images with high image quality. The processed images were simultaneously analyzed using the objective evaluation metrics Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). The findings demonstrate that the dehazing effect is favorable compared to other state-of-the-art dehazing algorithms, achieving a PSNR and SSIM of 14.7 and 0.54 for the indoor images, and 16.54 and 0.54 for the outdoor images respectively using the NTIRE 2018 dataset. Using the SOTS dataset, the model achieved a PSNR and SSIM of 23.98 and 0.87 for the indoor images, and 19.88 and 0.83 for the outdoor images.