Blind image quality assessment using natural scene statistics of stationary wavelet transform

Image quality assessment (IQA) by human observers, who are the ultimate users, is of utmost importance. Since IQA performed by human observers could be time consuming therefore, computational models are required that can assess the quality of images using objective metrics. Blind image quality assessment (BIQA) methods could predicts the perceived quality of distorted images without requiring information regarding the pristine version of the image. The natural scene statistics (NSS) based BIQA methods predict the quality score by extracting features either in spatial or transform domain. We propose a new BIQA technique that extracts features both in spatial and transform domain. The NSS of stationary wavelet transform (SWT), morphological gradient, and discrete Laplacian are known to be similar to human visual system, and hence can be used for BIQA. Morphological gradient has the advantage that it is easy to generalize for any type of feature space and discrete Laplacian has the property to retain key structural information inherent in the image. SWT provides edge and high frequency information while preserving the scale of the image. We remove the redundancies between the extracted features from spatial and transform domain using an adaptive joint normalization framework. Our proposed BIQA technique is tested on five publicly available databases (i.e., TID2008, TID2013, LIVE, CSIQ, and LIVE from the wild image quality challenge). Our experimental results have shown that our proposed BIQA technique achieves higher accuracy in terms of image quality prediction in comparison to the state-of-the-art BIQA and full reference IQA techniques.