MULTIPLE PATCH-AWARE NETWORK FOR FASTER REAL-WORLD IMAGE DEHAZING

This paper proposes a Multiple Patch-aware Dehazing Network (MPADN), to remove haze in real-world images fast and efficiently. Firstly, we design a Multiple Patch-aware Module (MPAM), which utilizes joint decisions from multiple patch awareness to gain more stable local features. Compared with the multi-patch hierarchical strategy, MPAM extremely compresses the size of the model. Besides, we propose a novel data enhancement method called Concentration Sampling Enhancement (CSE), which generates new training samples by haze concentration sampling based on hazy images and clear images. This method makes use of the precious limited real-world paired data to generate more potentially usable data. It is easy to extend CSE to more models or tasks where the ground truth and the input information are similar. Experiments show that MPADN outperforms state-of-art fast dehazing methods, and it only takes 0.0065s to process a 1600x1200 image.