A Novel Feature Set for Recognition of Printed Amazigh Text using Maximum Deviation and HMM

The growing need of Tifinagh characters recognition in several domains in Morocco such as education, telecommunication, etc, has made it a vital area of research. This paper presents a novel set of structural features generated on the Tifinagh character geometry. This set based on a vocabulary that consists of various fundamental strokes, which is generated using the intrinsic morphological characteristics of the Amazigh script. The input text image is undergoing several preprocessing operations: binarization, skew correction, line segmentation, character segmentation and size normalization. Indeed, the obtained isolated characters are first pre-classified into one of two character groups (circular, non-circular) using the Hough Transformation method. Then, each one is described with their points that have maximum deviation and their segments. Thereafter, each segment of the character is transformed into primitives sequence. We use the discriminating path (DP-HMM) recognition system witch operates on proposed vocabulary. Only one model is built and trained on all elements of this vocabulary. Each path through this trellis represents a sequence of segments, i. e. the character of the Tifinagh alphabet. Finally, the recognition is performed by dynamically decoding the optimal path according to the criterion of maximum likelihood. The obtained scores show the robustness of the proposed approach