NO-REFERENCE IMAGE QUALITY ASSESSMENT BASED ON DISTORTION SPECIFIC AND NATURAL SCENE STATISTICS BASED PARAMETERS: A HYBRID APPROACH

In web applications, for efficient use of bandwidth and storage requirement, images are compressed with lossy techniques. Quality of images degrades due to the compression artifacts. Transmission channels also contribute in degradation by adding noise. Image quality assessment plays a vital role to address this issue. In majority of cases, reference image remains unavailable for the assessment. Thus, No-Reference quality assessment techniques are widely used. In this paper, machine learning based hybrid approach for NR quality assessment is proposed. Blockiness based parameters, other statistical parameters and NSS based features are provided as input to the feed forward neural network. The back propagation training algorithm predicts a quality score. This score is correlated with differential mean opinion score (DMOS). Here, input can be images of any type from best to worst quality, and the approach used exploits the nonlinearity in the behaviors of parameters. It has been noticed that the predicted score correlates well with the DMOS with 93% accuracy. The above parameters are also used as input to Support Vector Machine for classification. The observed accuracy of this classifier is 89%.