Improving OCR quality of documents using generative adversarial networks

Description

While we move toward a digital world, paper-based document processing is still an integral part of business processes across a wide range of domains from finance to healthcare. Traditionally, these documents are processed manually and natural language processing (NLP)-based document processing solutions bring significant efficiencies in this manual process by automating document classification, extraction, or search. OCR is a crucial step in any document-processing task as it enables machine interpretation of text and parsing textual information. However, issues in document quality, such as blurred text, fragmented characters, merging characters, low resolution, skew, and noise, can tremendously affect OCR accuracy and degrade its performance. In particular, images are corrupted through stains, wrinkles, pixel noise, and impulse noise added during scanning. Poor resolution leads to the merging of character strokes with the document background.

Nagendra Shishodia, Solmaz Torabi, and Chaithanya Manda explore the general adversarial network (GAN) developed at EXL to overcome these challenges of enhancing the resolution and denoising scanned images. The model was trained using custom generated data combined with externally available datasets. Once trained, the generator’s able to increase the resolution of the document by a predefined factor, sharpen the character borders, increase contrast, and eliminate pixel noise while preserving edge features of characters. The model significantly improves OCR accuracy (using open source and commerical OCR systems) on a word and character level. Implementing the GAN model also helped the company achieve significant efficiencies in multiple document-processing solutions from finance to healthcare. The model also helped improve accuracy of handwriting detection APIs.

You’ll learn how GANs can bring efficiency and accuracy in document-processing pipelines. In particular, Nagendra, Solmaz, and Chaithanya detail how the dataset was created for this task, the model architecture and training methodology, and they showcase the performance of the model across multiple document-processing applications and examine potential next steps.