Improved optical character recognition with deep learning

Optical Character Recognition (OCR) plays an important role in the retrieval of information from pixel-based images to searchable and machine-editable text formats. For instance, OCR is typically used in many computer vision applications such as in automatic signboard recognition, language translation as well as in the process of digitizing scanned documents. However, compared to old documents or poorly printed documents, printed characters are typically broken and blurred, which makes the character recognition in potentially far more complicated. Although there are several OCR applications which utilizes techniques such as feature extraction and template matching for recognition, these methods are still not accurate enough for recognition. In this work, deep learning network (transfer learning with Inception V3 model) is used to train and perform OCR. Deep learning network is implemented and trained using Tensorflow Python API that supports Python 3.5+ (GPU version) which is available under the Apache 2.0 open source license. The Inception V3 network is trained with 53,342 character images consisting of noises which are collected from receipts and newspapers. From the experiment results, the system achieved significantly better recognition accuracy on poor quality of text character level and resulted in an overall 21.5% reduction in error rate as compared to existing OCRs. Besides, there is another experiment conducted to further analyze the root causes of text recognition failure and a solution to overcome the problem is also proposed. Analysis and discussion were also made on how the different layer’s properties of neural network affects the OCR’s performance and training time. The proposed deep learning based OCR has shown better accuracy than conventional methods of OCR and has the potential to overcome recognition issue on poor quality of text character