A set-based benchmarking method for address bloc location on arbitrarily complex grey level images

摘要

A set-based benchmarking method has been previously presented for page segmentation. The author proposes to adapt it to address bloc location on grey level images of various layouts and visual quality. For each mailpiece, a ground truth file describing the address bloc lines is compared with the automatically located zones meant to contain them. The pixels forming the ground truth lines are extracted by an adaptive binarization on the manually located zones. The algorithm allows ABL to locate lines partially, to provide several solutions, and the contents-related importance of a text line is reflected numerically. It provides an explicit error diagnosis together with a numerical evaluation.