Natural Image Statistical Significant Graph Model And Quality Blind Assessment

Images of the visual environment captured using high-quality capture devices operating in the visual spectrum are broadly classified as natural images, including images of outdoor scenes, manmade objects and indoor environments, which differentiates them from text, computer generated graphics, paintings and drawings, random noise, or images captured from nonvisual stimuli such as radar and sonar, X-rays, ultrasounds etc. Thus, natural scenes form an extremely tiny subset of the set of all possible images. The prior knowledge of this subset as well as the corresponding statistics models forms the basis of many application of the fields of image processing and computer vision.In this paper, a natural scene statistical saliency map model is proposed based on natural scene statistics model, thus simulating the selective visual attention mechanism in human visual system. Based on the Gaussian scale mixture distribution of natural scene statistics, the multiplier variable is extracted to model the saliency map of natural images. Further, the saliency map model is extended to color images following the processing mechanism of chromatic information in human visual system. Our analytical as well as the experimental results show that the proposed saliency map model here is consistent with the visual attention mechanism of the human eyes. Namely, the presented saliency map model can highlight the visual stimulus with great saliency while suppressing those with repeat occurrences. Thus it can describe the saliency distributions of the visual stimulus of an image much better.Meanwhile, a statistical measure for blind image quality assessment (IQA) based on the natural scene statistics is proposed. The wavelet coefficients’distribution parameters of the distorted images are estimated based on the natural scene statistics and the image distortion model. The mutual information between the distorted and the corresponding reference images is further calculated from the estimated distribution parameters, thus quantifying the information fidelity as an efficient image quality assessment criterion. The proposed statistical measure does not require any prior information of the reference image and avoids the feature extraction and machine learning processes required by existing image quality blind assessment methods. Evaluated on the open IQA database, it is demonstrated that the proposed statistical measure corresponds well with the subjective human evaluations and outperforms the state-of-art blind IQA algorithms. Furthermore, saliency-based image quality assessment methods are introduced. By the comparison of the improvements brought by incorporating different saliency maps into the full-reference image quality assessment metrics, we verify the superiority of the natural scene statistical saliency map model over other saliency models. Particularly, the proposed natural scene statistical saliency map is incorporated into the proposed statistical measure for blind image quality assessment. Based on the evaluation results on the open IQA database, great improvement is shown over the existing blind IQA methods.