IW-GAN: Inverse-Wavelet GAN For Haze And Rain Removal

Abstract Rain and haze removal presents a significant challenge in computer vision. Despite their co-occurrence in natural environments, research addressing their simultaneous removal remains limited. This paper proposes a novel Inverse Wavelet Generative Adversarial Network (IW-GAN), that employs complex convolutional layers that allows the network to learn dehazing and deraining during the inversion process. To optimise model performance, we introduce a custom loss function that combines GAN loss, L1 loss, and Multi-Scale Structural Similarity Index (MSSIM) loss. Additionally, we extend our analysis beyond the Haar wavelet family, training the proposed model using various wavelet families, including Daubechies, Symlets, Coiflets, and Meyer, across four datasets: RainDID, Rain800, and RESIDE6K ITS and OTS. Results demonstrate strong performance of IW-GAN over existing methods in haze and rain removal tasks. This highlights the potential of our approach for practical applications in improving image clarity under adverse weather conditions.