NEW DISCREPANCY MEASURE FOR EVALUATION OF SEGMENTATION QUALITY

摘要

A new quality measure is proposed for evaluating the performance of segmentation algorithms. This discrepancy method is based on object-by-object comparison of a segmented image (machine segmentation) versus ground-truth segmentation (reference image). The error measure we propose offers two main advantages compared to other quality measures. The first advantage to mention is the use of alternative methods in computing the distance from the contour of the segmented object to the reference one. This method takes into consideration the interior of the object and eliminates the inconveniences that appear in the case of the concave objects. The second improvement comes from adding a weighted shape fitting score: the score of the segmented contour is enhanced by a factor which indicates the similarity between these two curves.