Performance measurement without ground truth to achieve optimal edge

摘要

This paper presents an effort to develop performance measurement of edge detector without ground truth. The motivation is to drive edge detector to catch up with human vision that is capable of always generating closed-loop edge. Two axioms of edge are presented, i.e. the presentation of edge as a closed-loop line and its characteristic as being thin. A set of features and approaches are generated in this research in order to enable the measurement of edge map. Experiment shows that current edge detector is still difficult to catch up with human vision for recognising edge. Experiment also proves that statistical approach based on object size delivers the best performance measurement among other approaches. And, the use of region of interest is proved useful to optimise edge measurement.