Design considerations for image segmentation quality assessment measures

Abstract

Factors to consider when designing quality assessment measures for image segmentation are discussed. Quality assessment requires one manually generated segmentation (for reference) plus computer-generated segmentations corresponding to different image segmentation algorithms or algorithm parameter settings. Since true pixel class assignments are seldom available, one must typically rely on a trained human analyst to produce a reference by using a mouse to draw boundaries of perceived regions on a digital image background. Different algorithms and parameter settings can be compared by ranking computed disparities between maps of computer-generated region boundaries and region boundaries from a common reference. Proximity-based association between two boundary pixels is discussed in the context of association distance. Motivated by the concept of phase-modulated signals, a penalty factor on the degree of association is then introduced as some non-negative power (phase modulation order) of the cosine of disparity in phase (boundary direction) between two boundary pixels. Families of matching measures between maps of region boundaries are defined as functions of associations between many pairs of boundary pixels. The measures are characterized as one-way (reflecting relationships in one direction between region boundaries from two segmentations) vs. two-way (reflecting relationships in both directions). Measures of inconsistency between perceived and computed matches of computer and manually generated region boundaries are developed and exercised so that effects of association distance, phase modulation, and choice of matching measure on image segmentation quality assessment can be quantified. It is quantitatively established that consistency can be significantly improved by using two-way measures in conjunction with high-order phase modulation and moderate association distances.