Optical Character Recognition (OCR) in Cursive Scripts Using Object Detection Networks

Document Type : Original Article

Authors

Jundi-Shapur University of Technology, Dezful, Iran.

Abstract

Optical Character Recognition (OCR) in cursive scripts, where the letters of a word are joined in a flowing manner and overlap in both directions, deals with the struggles raised while segmentation of unrecognized characters and recognition of unseparated characters. In this paper, we propose using object detection models for character detection in cursive scripts. Simplicity of implementation and efficiency of this method in recognition of handwriting-style fonts are investigated and discussed. Here, YOLO model is used to separate and classify the characters of arbitrary three-letter words in Persian script as a case study. Initially, we generated synthetic datasets suitable for the YOLO network from handwriting-style Persian fonts, such as Maneli and IranNastaliq. By using the YOLO model, we achieved high Precision of 98.5% in character detection of Maneli font and 97.6% for a mixture of words in Maneli and IranNastaliq fonts, while the accuracy for the regular font Arial was almost 100%. Then, we challenged the proposed model by adding noise, blur, and skewness to the samples. Furthermore, we utilized a multi-layer perceptron (MLP) model to predict the words from the characters detected and localized by YOLO with the accuracy of 99.8% for Maneli font and 97.7% for a mixture of words in Maneli and IranNastaliq fonts, while the word detection accuracy for the regular font Arial was almost 100%. This approach enables us to recognize complete words accurately from complex handwriting-style fonts, without using a Persian vocabulary dictionary.

Keywords

Main Subjects

TABRIZ JOURNAL OF ELECTRICAL ENGINEERING

Articles in Press, Corrected Proof
Available Online from 04 December 2024

Files

Share

How to cite

Statistics