A confidence prior for image dehazing. (November 2021)

• Record Type: Journal Article
• Authors: Yuan, Feiniu; Zhou, Yu; Xia, Xue; Qian, Xueming; Huang, Jian
• Is Part Of: Pattern recognition. Volume 119(2021)
• Publication Date: 2021-11
• Subjects: Regression -- Classification -- Image dehazing -- Confidence prior -- Appearance feature; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108076
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a unified framework for better explanation of several existing priors. Under the unified framework, we derive a confidence prior that uses a ratio to freely adjust the removal degree of outliers or noises. To solve heterogeneity of image signals and abrupt depth jumps in hazy images, we use a learning method to adaptively estimate a confidence ratio for each pixel. Abstract: By sorting channel-minimized values in an ascending order, we individually put the values of several existing image dehazing priors on the curve of sorted values to propose a framework for unifying and understanding these priors. Then we propose a confidence ratio to specify the probability of each channel-minimized value within a range, and thus we can intuitively find a suitable point from the curve, which is actually defined as a novel prior. Although our novel prior and existing ones are perfectly unified under the same framework, our prior has an important advantage that it can freely control the suppression degree of outliers by directly adjusting the confidence ratio of channel-minimized values. In this way, we can remove influence of outliers in a controllable manner. To solve the problems caused by heterogeneity of pixel values and abrupt jumps of scene depths in hazy images, we adopt a regression method to adaptively learn the relationship between patch appearance and confidence ratios for all pixels. To further improve robustness, we use a Gaussian kernel to smooth the estimated confidence ratios for local consistency. Extensive experiments on both natural and synthetic images show that our confidence prior achieves significantly better performance than existing state-of-the-art methods.