No-Reference Image Quality Assessment for Defocus Restoration

Even though both full-reference and no-reference image quality assessments are extensively applied to evaluate degraded images e.g. blurred, noised, compressed and distorted, few of them especially the no-reference species are suitable for image restoration. The goal of this work is to propose a no-reference quality assessment for defocus restoration. Because of that most optical systems have circular aperture, the defocusing is an isotropic filter to blur the edges. As a result, the change of edge location detected by Canny algorithm will be small after defocusing. Otherwise, smooth regions will be smoother after defocusing. Therefore, the locations of edges and smooth regions will not change. We get those invariant points from defocus image to form two point sets, one containing edge points, another containing the centers of smooth regions. The two point sets are used to form patches in restoration image, and image patch structure metric is proposed for those patches to score the local restoration quality. At last, the quality assessment for defocus restoration is modeled as an overall measure by all of the patches structure metric. Experiments on simulated defocus image and real defocus images are presented to demonstrate the effectiveness of this assessment.