Pixel-Level Multidirectional Image Sharpness Linear Assessment for Optical Image Stabilizer Performance Monitoring

Abstract

Linear quantification of blur intensity in different directions of an image and blur-type classification are crucial for the detection of anomalies and optimization calibration of widely used optical image stabilizer (OIS). However, previous image quality assessment methods mainly focused on isotropic blur, emphasizing the correlation with subjective ratings, making it difficult to provide multidirectional linear sharpness assessment. Moreover, they invariably offer a single evaluation score regardless of the kind of fuzziness. In this article, we propose a pixel-level multidirectional image sharpness (PMIS) linear assessment method that, to our knowledge, is the first to provide the capability to linearly quantify blur across multiple directions and distinguish blur types in an end-to-end manner while assessing the degree of image blur. We introduce a novel method for extracting edge information to characterize image blur, significantly enhancing correlation with the human visual system (HVS) in blur perception by filtering out high-frequency edge noise. Uniform edge block selection and data postprocessing are introduced to adapt to HVS characteristics and enhance robustness. Using blur results from four different directions and simply setting thresholds, we are able to achieve an 84.5% classification accuracy in distinguishing between defocus blur, motion blur, and clear images. In addition, we creatively make a motion blur image quality (MBIQ) database, accurately representing motion blur through the concrete physical quantity of rotational speed. Experimental results confirm that PMIS achieves significant improvements over the previous methods especially on databases with anisotropic motion blur.