Image dehazing aims to remove the haze noise and restore the image content from hazy images. It is a challenging task because of the unbalanced distribution of the haze noise and the variety of the image contents. Most existing methods apply convolutional neural networks to learn the dehazing process by blind end-to-end training, which relies on the noise distribution of the training datasets. Indeed, the distribution of the haze noise is much related to the image depth, which indicates the distances of the scenes from the viewer. In this work, based on the structure of conventional generative adversarial network, we propose a depth aware method to estimate the depth maps and provide the depth features for dehazing within one joint framework. By fusing the depth feature to the dehazing network, the dehazing model is able to better separate various extent of haze against the image content. The experiments demonstrate that the proposed depth aware module can significantly improve the performance of the dehazing model and is able to be implemented into traditional CNN-based dehazing models conveniently.
- Depth aware
- Feature fusion
- Generative adversarial network
- Image dehazing
ASJC Scopus subject areas
- Computer Vision and Pattern Recognition
- Computer Graphics and Computer-Aided Design