On Verification of Blur and Sharpness Metrics for No-reference Image Visual Quality Assessment

Abstract

Natural images may contain regions with different levels of blur affecting image visual quality. No-reference image visual quality metrics should be able to effectively evaluate both blur and sharpness levels on a given image. In this paper, we propose a large image database BlurSet to verify this ability. BlurSet contains 5000 grayscale images of size 128×128 pixels with different levels of Gaussian blur and unsharp mask. For each image, a scalar value indicating the level of blur and the level of sharpness is provided. Several image quality assessment criteria are presented to evaluate how a given metric can estimate the level of blur/sharpness on BlurSet. An extensive comparative analysis of different no-reference metrics is carried out. Reachable levels of the quality criteria are evaluated using the proposed blur/sharpness convolutional neural network (BSCNN).