A novel sparsity-inspired blind image quality assessment algorithm

We present a novel blind image quality assessment (BIQA) algorithm inspired by the sparse representation of natural images in the human visual system (HVS). The hypothesis behind the proposed method is that the properties of natural images that afford their sparse representation are altered in the presence of distortion. We attempt to quantify this change in sparsity and show that it is indeed a measure of the unnatu-ralness or distortion in an image. We first construct an over-complete dictionary from a set of pristine images using the K-SVD algorithm. This dictionary is then used to sparsely represent a different and significantly smaller set of pristine images to extract "reference" features. To evaluate the quality of a given image, features are extracted from its sparse representation and quantified with respect to the "reference" features. We call our algorithm Sparsity-based Blind Image Quality Evaluation (SBIQE). We show that the proposed algorithm consistently correlates well with subjective scores over several popular image databases. Further, it compares reasonably with state-of-the-art BIQA algorithms. Additionally, our algorithm is both opinion-unaware and distortion-unaware.