Dark channel map and union training strategy for object detection in foggy scenes

Abstract

Most existing object detection methods in real-world hazy scenarios fail to handle the heterogeneous haze and treat clear images and hazy images as adversarial while ignoring the latent information beneficial in clear images for detection, resulting in sub-optimal performance. To alleviate the above problems, we propose a new dark channel map-guided detection paradigm (DG-Net) in an end-to-end manner and provide an interpretable idea for object detection in hazy scenes from an entirely new perspective. Specifically, we design a unique dark channel map-guided feature fusion (DGFF) module to handle the adverse impact of the heterogeneous haze, which enables the model to focus on potential regions that may contain detection objects adaptively, assign higher weights to these regions, and thus improve the network’s ability to learn and represent the features of hazy images. To more effectively utilize the latent features of clear images, we propose a new simple but effective union training strategy (UTS) that considers the clear images as a complement to the hazy images, which enables the DGFF module to work better. In addition, we introduce Focal loss and Self-calibrated convolutions to enhance the performance of the DG-Net. Extensive experiments show that DG-Net outperforms the state-of-the-art detection methods quantitatively and qualitatively, especially in real-world hazy datasets.