A Statistical Frame Work to Identify Type and Level of Image Distortion

Abstract

Growth and usage of digital images have been tremendous in recent years as they are the source for representing and communicating information. Various algorithms are developed to improve the performance of an image when it is subjected to distortions viz., acquisition, transmission, compression. It is highly undesirable to evaluate the quality score or quantitative index based on Human Visual System (HVS) in practical systems. A large number of restoration algorithms are available to enhance quality. So, a quality metric needs to be deployed to determine which algorithm automatically predicts the quality score of an image based on the availability of reference(original) image and without availability, and it should provide the best results. Blind Image Quality Assessment (BIQA) aims to identify the distortion level and type blindly without prior knowledge about the reference image. Contrast features convey most of the skeletal information about an image. Work is to extract the features based on the joint statistics of adaptive normalization of Gradient Magnitude (GM) and Laplacian of Gaussian (LOG). These normalized responses are quantized into different levels and Q, respectively. For the quantized levels of GM and LOG, estimate the Marginal and conditional probabilities, i.e., P _GM and P _log, c _g and c _l. When an image is unnatural, the shapes of P _GM and P _log, c _g and c _l will be different when compared to the natural image. To estimate the quality score, the correlation between probabilities using Spearman and Pearson methods is used and the classical tool SSIM. The Proposed model is evaluated on the CSIQ 2013 database for different levels of and Q, the correlation coefficients will follow specific relationship to identify the distortion type, and its performance is highly competitive when compared with some of the NR-IQA algorithms.