ABSTRACT
In this paper, Bangla handwritten compound character recognition by using simple yet robust Local Directional Pattern (LDP) features is studied and reported. LDP is robust as it is gradient based and resistant to non-monotonic illumination changes. No single algorithm can recognize the alphabets of different script languages. Besides, Bangla script contains a higher order of alphabet set, and recognition of these alphabets with high accuracy is a challenging task. Here, the character images are divided into blocks and sub-blocks based on the Centre of Gravity (COG) and features are extracted from each block. A total of 1176 dimension feature vectors are extracted. The Bangla data set, CMATERdb3, is used here in the experiment by considering 50 randomly selected character classes. In each class, 100 samples are taken and 2390 images are tested. Experimentally, 98.8% average accuracy is achieved by using LDP feature.
