Image Quality Assessment Based On SIFT Feature And Structural Similarity

Artifacts or errors may be introduced into the digital image in the process of compression, transmission, storage and display, etc. Image quality assessment would be used to evaluate the effect of different image processing algorithms, providing reference for optimizing image processing algorithms. An effective algorithm for image quality assessment is significant in the field of image processing. So, in this article, the method, combining the SIFT feature and the structural similarity, is explored to evaluate the image quality.SIFT(Scale Invariant Feature Transform) is a kind of local feature description, which is invariable to image translation, scaling, rotation and widely used in image matching, retrieval as well as video processing, etc. SIFT features is highly robust to the affine transformation, viewpoint change, illumination change and noise. By Gaussian smoothing and down-sampling, Gaussian pyramid of image is established and the SIFT feature is extracted based on the multi-scale space; By exhaustion method, SIFT feature matching between the original image and the targeting image is executed.SSIM (Structural Similarity Index Measure) algorithm is studied in this article. SSIM algorithm takes account of the contrast, illumination and structure between images in some specific areas, evaluating the similarity from the three aspects independently, at last getting the global similarity by accumulating the similarity of each pairs. On the basis of that, SSIM algorithm is expanded to multi-scale space, and pixels match reflecting the structural information is built in Gaussian pyramid.SIFT feature pixels matching and SSIM algorithm are combined to research an objective quality assessment simulating the human visual system. Based on the multi-scale space, SIFT feature pixels matching between the original image and the retargeted image is established from top to bottom level. At the same time, local non-feature pixel pairs within two images are built based on structural similarity. The match of pixel pairs are delivering from the coarse level to the exact level in Gaussian pyramid. Pixels match between the images are built from the global to the local structure. Considering the different importance of pixels due to the whole image quality, fetching the weighting to calculate the global similarity by introducing the saliency maps.Selecting the CIELab color model to process all kinds of color images. Subjective assessment experiment is made to verity the effectiveness of our algorithm. Experimental results show that our algorithm can assess the quality of the retargeted image quantitatively, which tends to the results of subjective assessment experiment.