Intelligent Character Recognition with Shared Features from Resnet

Abstract

Transcribing handwritten documents is still one of the hardest problems in the field of computer vision. Digitizing handwritten documents have many potential applications. Even today, people do this task manually. Online handwritten recognition, takes advantage of the temporal information, and transcription is easy compared to offline handwritten recognition, where temporal information is absent. Also, there are numerous ways people write, and the handwriting of one person varies from that of others. Generalizing this makes handwritten text recognition much difficult. In this paper, we proposed a single neural network for both word localization and word recognition. Our proposed intelligent character recognition (ICR) method has Resnet architecture to extract features from a handwritten document. These shared features are used for both localization and prediction. To overcome the scaling issues in localization, we considered feature pyramidal networks (FPN). Later the word predictions are corrected with the help of dictionary based on the Levenshtein Distance (LD). Since there are not enough handwritten documents available, we built our own dataset from answer sheets of local students. Our ICR is trained and tested on three datasets. Across the three datasets, we obtained an average character error rate (CER) of 12.06% and a mean Average Precision (mAP) of 0.94.