SVM with inverse fringe as feature for improving accuracy of telugu OCR systems

Abstract

Designing an OCR system with high accuracy is quite a tough task as the system performance gets affected by its component modules. The accuracy and quality of the OCR system depends on impact of each module. The overall system performance changes if there is an improvement in a module. In our work at present, we have developed an OCR system for Telugu (Drishti System). We proposed in our paper SVM algorithm with inverse fringe as feature for Telugu OCR. The idea is to improve the performance of system by increasing recognition accuracy of the developed system. Support vector machines (SVM) was shown by several researchers to deliver high performance on Indic OCRs. SVMs have been applied to Telugu OCR and are tested with different features. In our experiments, we used fringe distance and its complementary version, the inverse fringe as a feature to the SVM. These two features have been used to develop the working model of Telugu OCR with an accuracy approaching 90%. It is shown that the performance is good over more than 300 classes. With inverse fringe as feature, the system with 325 classes is trained with 15543 labeled Telugu characters and tested over 75335 unlabeled Telugu characters; the accuracy of the system is found 99.50%. The SVM-based classifier is tested on our scanned image document corpus of more than 4500 pages and about 5,000,000 symbols. Evaluation of end-to-end system performance is done in our experiments. From the results, it has been depicted that SVM classifier is giving an improvement of approximately 1.24% over the developed Telugu OCR (Drishti System).