Subset-Guided Consistency Enhancement Assessment Criterion for an Imageset Without Reference

Abstract

In a new era for machine vision, some image enhancement algorithms can be used to improve the quality of an imageset without reference. To assess the performance of imageset enhancement, the existing average criterion utilizes a no-reference image quality metric to calculate the quality score of each enhanced image, and quantifies the performance of an enhancement algorithm by the mean value of all scores. If the quality scores of some images fluctuate greatly, their mean value difficultly reflects the degradation or worst cases during imageset enhancement. Therefore, this paper analyzes and illustrates the need and significance of consistency enhancement assessment, and then proposes a subset-guided consistency enhancement assessment criterion for an imageset without reference. By measuring the subset of an imageset, the proposed criterion firstly calculates the difference of quality scores of each image before and after enhancement and then filters the outlier data outside confidence interval, and finally quantifies the consistency enhancement performance of an enhancement algorithm according to its consistency enhancement degree. When a small subset is used to guide its large imageset, the average criterion judges a consistency or non-consistency enhancement algorithm with a 16.7% false identification ratio, and also makes one misjudgment about the optimal-consistency algorithm, while the proposed criterion always correctly judges the non-consistency or optimal-consistency enhancement algorithm. This paper can help the scientific community to select a robust enhancement algorithm in the degradation or worst cases. As compared with the average criterion, the proposed criterion is more robust in terms of subset-guided consistency enhancement assessment, which may effectively find an optimal-consistency or non-consistency enhancement algorithm for the rest of an imageset.