Classification for Low Intra-Class Variability and High Number of Classes

Abstract

Optical Character Recognition (OCR) is the recognition of handwritten or printed text for various digital processing tasks. It is an important area of research in the field of image processing, natural language processing and artificial intelligence. Many real world applications, for example price tag scanners, online translators and text to speech converters are based on OCRs. The research and work for English OCR has been quite remarkable. Many robust OCR systems have been developed for English language. The research on Urdu OCR is quite recent and till date, no sophisticated Urdu OCR system exists. Urdu is the national language of Pakistan and is spoken by over 300 million people around the world. Owing to this importance, there is a need to develop a method for recognition of Urdu script. The lack of attention to Urdu OCR is due to the complexity of this language. It is a highly cursive and context sensitive language. A single word has numerous struc tural variations, which make it difficult to be recognized. Due to these challenges, no benchmark dataset for Urdu could be developed. The focus of our research is to develop an effective method for recognition of Urdu script. In our proposed method, we developed an Urdu ligature dataset named CEFAR dataset and used deep learning to recognize these ligatures. The dataset contained exhaustive com binations of Urdu characters of length 2 and 3. These ligatures were in the form of images divided into 3 parts; 2 and 3 character ligature sets separately and the third part contained ligatures of both 2 and 3 characters. The aim was to train a deep learning model for recognizing these ligatures. The main challenge associ ated with this data was its high number of classes and low intra-class variability. In our proposed method, we developed a novel technique to solve this problem by using data augmentation to increase the number of representative samples within each ligature class and then applied class redefinition for class reduction. Data augmentation was followed by ligature recognition using Recurrent Neural Net works. RNNs were employed for classification of these ligatures. A 34 layer recurrent neural network model was used with a final fully connected layer for classification. Three separate models were trained and evaluated for the three lig ature sets. All the three ligature’s training models gave remarkable performance. The 2 character ligature set gave an accuracy of 96.4%, 99.7% accuracy on three character ligature set and 97.5% accuracy on the combined ligature set of 2 and 3 characters. The overall performance considered for the recognition system was the accuracy of the model over the combined 2 and 3 character ligature dataset, which was 97.5%. Our model performed brilliantly well than the existing deep learning methods for Urdu ligature recognition. The excellent classification accu racy of our deep learning model makes this research play an effective role in not only building a benchmark Urdu dataset but also its classification. In future, this research could be extended for development of classification systems for ligatures of lengths greater than three.