Single Image Dehazing Using Ranking Convolutional Neural Network

Abstract

Single image dehazing, which aims to recover the clear image solely from an\ninput hazy or foggy image, is a challenging ill-posed problem. Analysing\nexisting approaches, the common key step is to estimate the haze density of\neach pixel. To this end, various approaches often heuristically designed\nhaze-relevant features. Several recent works also automatically learn the\nfeatures via directly exploiting Convolutional Neural Networks (CNN). However,\nit may be insufficient to fully capture the intrinsic attributes of hazy\nimages. To obtain effective features for single image dehazing, this paper\npresents a novel Ranking Convolutional Neural Network (Ranking-CNN). In\nRanking-CNN, a novel ranking layer is proposed to extend the structure of CNN\nso that the statistical and structural attributes of hazy images can be\nsimultaneously captured. By training Ranking-CNN in a well-designed manner,\npowerful haze-relevant features can be automatically learned from massive hazy\nimage patches. Based on these features, haze can be effectively removed by\nusing a haze density prediction model trained through the random forest\nregression. Experimental results show that our approach outperforms several\nprevious dehazing approaches on synthetic and real-world benchmark images.\nComprehensive analyses are also conducted to interpret the proposed Ranking-CNN\nfrom both the theoretical and experimental aspects.\n