Rule based segmentation of lower modifiers in complex Bangla scripts

Segmentation is the most challenging part of Bangla optical character recognition (OCR). To solve the problems of joining errors, several algorithms have been proposed in the literature, with varying degrees of accuracy. The selection of the lower modifier container units and the subsequent extraction of the modifiers from the core unit during segmentation have not been studied extensively. We present a dissection based lower modifier segmentation method which solves the problem of segmenting lower modifiers under a wide range of document images. A key goal in our methodology is to avoid over-segmentation of the units that do not actually contain any lower modifier, leading to unacceptably high error rates during segmentation. Our methodology consists of four tasks: we first identify the lower modifier separator line using character height information, and then select the primary lower modifier containers; we filter this set to eliminate the units/characters that do not actually contain any lower modifier; we then extract the lower modifier unit using the features of the core units and the lower modifiers; the final step consists of a set of empirical rules, aided by dictionary lookups, to eliminate most of the errors, resulting in an accuracy of 99.6%.