A Retinex Prior to Multi-Scale Fusion for Single Image Dehazing

Abstract

This Image-Dehazing paper proposes to combine the Multi-Scale Fusion technique with the Retinex Algorithm. The paper proposes to extract reflectance matrices and incorporate them into the multi-scale fusion algorithm. The technique proposed aims to reduce the halo effect observed in image-dehazing applications and related works for heavily hazy images. Moreover, an improvement in the quality of the output using the proposed novel algorithm is observed. Quantitative, as well as a visual display of results, using the DENSE HAZE dataset, give an accurate interpretation of the effectiveness of the proposed work. The best value of Structural Similarity Index (SSIM) obtained is 0.9128 which shows a 62% increase in image quality as compared to average SSIM values of previously known methods. The Peak Signal to Noise Ratio (PSNR) and Mean Square Error (MSE) show improvement by 78% (TT Playroom) and 95% (Castle) respectively. To allow analysis with regards to pixel compression that may have resulted during the process, two No Reference Image Quality Metrics have been also computed.