Mobile-UNet GAN: A single-image dehazing model

Dehazing, a crucial task in computer vision applications, aims to remove atmospheric haze from images. The emergence of generative adversarial networks (GANs) as a powerful solution for image restoration, including dehazing, has garnered recent attention. In this paper, an innovative approach for dehazing is introduced, utilizing a GAN based on the UNet architecture. To ensure effective extraction and encoding of vital features from hazy images, the model incorporates various encoder blocks, namely modified-MobileNet, EfficientNet, ResNet, and VGG19. Through extensive experimentation and evaluation, it is revealed that the modified-MobileNet encoder block proves to be the most efficient for the dehazing GAN. The outcomes of the experiments and comparison analyses provide strong evidence supporting the efficacy of the suggested approach in achieving high-quality dehazing results. The proposed method achieves less run-time than the existing models with run-time of 0.1 image/s.