Towards Domain Invariant Single Image Dehazing

Presence of haze in images obscures underlying information, which is undesirable applications requiring accurate environment information. To recover such an image, a dehazing algorithm should localize and affected regions while ensuring consistency between recovered its neighboring regions. However owing to fixed receptive field convolutional kernels non uniform distribution, assuring difficult. In this paper, we utilize encoder-decoder based network architecture perform the task integrate spatially aware channel attention mechanism enhance features interest beyond traditional conventional kernels. ensure performance across diverse range densities, greedy localized data augmentation mechanism. Synthetic datasets are typically used large amount paired training samples, however methodology generate samples introduces gap them real accounting for only distribution overlooking more realistic scenario non-uniform resulting inferior when evaluated on datasets. Despite this, abundance within synthetic cannot be ignored. Thus datasets, train proposed adversarial prior-guided framework that relies generated image along with low high frequency components determine if properties dehazed matches those ground truth. We preform extensive experiments validate domain invariance domains report state-of-the-art (SoTA) results. The source code pretrained models will available at https://github.com/PS06/DIDH.