Image dehazing via self-supervised depth guidance

Yudong Liang, Shaoji Li, De Cheng, Wenjian Wang, Deyu Li, Jiye Liang

Abstract: Highlights•We build a self-supervised image dehazing framework with self-supervised depth guidance, which sequentially generates hazy inputs, estimates the depth for hazy images with the aid of the depth estimations from clear images, and exploits the interactions between depth and hazes for image dehazing.•We design depth-guided hybrid attention Transformer blocks to exploit the correlations between the image depth and hazy densities in the images, which adaptively leverage both the cross-attention and self-attention to effectively model hazy densities via cross-modality fusion and capture global context information for better feature representations.•Our method obtains the state-of-the-art dehazing performances compared with the unsupervised or self-supervised dehazing methods. The self-supervised depth estimations not only improve the model generalization ability across different dehazing datasets, but also benefit the downstream detection tasks on hazy images.