Multi-Scale Adaptive Dehazing Network

Since haze degrades an image including contrast decreasing and color lost, which has a negative effect on the subsequent object detection and recognition. single image dehazing is a challenging visual task. Most existing dehazing methods are not robust to uneven haze. In this paper, we developed an adaptive distillation network to solve the dehaze problem with non-uniform haze, which does not rely on the physical scattering model. The proposed model consists of two parts: an adaptive distillation module and a multi-scale enhancing module. The adaptive distillation block reassigns the channel feature response via adaptively weighting the input maps. And then the important feature maps are separated from the trivial for further focused learning. After that, a multi-scale enhancing module containing two pyramid downsampling layers is employed to fuse the context features for haze-free images restoration in a coarse-to-fine way. Extensive experimental results on synthetic and real datasets demonstrates that the proposed approach outperforms the state-of-the-arts in both quantitative and qualitative evaluations.

1. Introduction

In poor weather conditions, the digital image will be degraded when it's photographed in the hazy scene, because the particles such as the haze and dusts interfere with the imaging process which results in the low contrast, color distortion and blurring problems. Because of this, dehazing which can improve the quality of hazy images is highly desired in photography, autonomous vehicles and other diversity computer vision applications [16].