Local features-based evidence glossary for generic recognition of handwritten characters.

Recognition of handwritten characters has been a challenging task so far. There exist thousands of official languages across the globe which are used to communicate through documentation. Optical character recognition (OCR) being the challenging domain in such context where images of such documents are to be recognized either offline or online. Online and offline recognition of documents refer to the approaches whereby the basic operation of character recognition is performed during documentation itself and recognizing characters from stored documents, respectively. Numerous applications from a range of fields like medical transcription, digitization of ancient manuscripts, language translations, etc. are solely dependent on the task of OCR. In this work, an efficient framework is presented for the purpose of handwritten character recognition that can be well utilized for both offline and online processes. The proposed work takes the handwritten character images as input. It applies set of pre-processing such that the samples become suitable for the feature extraction task. The novelty lies in the process of feature extraction whereby three distinct types of feature are considered based on the shape primitives of the images. These features are global to the sample. Subsequently, local shape features are further extracted out of this shape features after suitable quantization process. These local features are the evidences which can be generically used to recognize test samples. These local feature vectors are dubbed as evidences and are preserved into a glossary dubbed as evidence glossary. The efficiency of the proposed scheme is well justified as it utilizes only a fraction of the feature vector and still it can recognize the characters. Other advantages of the proposed work are scale invariance and rotations invariance. Suitable datasets from two distinct languages, namely Odia and English are utilized for evaluating the efficiency of the framework. Comparison of the performance of the framework with six distinct state-of-the-art machine learning models is conducted whereby it outclass the competent in terms of several performance metrics. [ABSTRACT FROM AUTHOR]