Entropy based single image dehazing with refined transmission using holistic edges

• Published in: Multimedia tools and applications Vol. 81; no. 14; pp. 20229 - 20253
• Main Authors: Lakshmi, T.R.Vijaya, Reddy, Ch.Venkata Krishna, Padmavathi, K., Swaraja, K., Meenakshi, K.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2022; Springer Nature B.V
• Subjects: Atmospheric models; Atmospheric scattering; Color; Computer Communication Networks; Computer Science; Darkness; Data Structures and Information Theory; Entropy; Fog; Haze; Image transmission; Light; Light sources; Multimedia Information Systems; Regression models; S parameters; Scattering coefficient; Special Purpose and Application-Based Systems; Visibility; Scattering coefficient; Hue disparity; Depth and Transmission maps; Non-achromatic atmospheric light; Haze removal
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-12485-z

The natural occurrences of haze, mist and fog obscures the optically captured outdoor scenes. The essential parameters of the atmospheric scattering model for dehazing are air-light (which depends on atmospheric light) and transmission map. Inaccurate estimation of these parameters leads to halo-artifacts, color distortions, etc. The performance of the available single-image dehazing models is limited by the color shifts caused due to the offset of light sources. The proposed work estimates the atmospheric light component by considering the lowest entropy from the quad-decomposed image as the haze opaque regions in the scene considered has low entropy. The transmission map is estimated by computing the scattering parameter, further refined with the holistic edges to calculate haze at different densities. A regression model is trained with the haze relevant features such as hue disparity, contrast, and darkness to compute the scattering coefficient rigorously. This improves the color shifts and visibility of the degraded outdoor scenes and mitigates the imbalances in the haze density concentrations in the scenes. The current model is evaluated using reference-based and non-reference based metrics to emphasize its perceptibility on hazy images when blended with pure light and non-achromatic light. The evaluations on real-hazy images signify that the true colors are well preserved, and the degree of visibility is improved compared to the state-of-the-art models.