Cloud-based simulation of page curling in the copying of documents
A Deep Neural Network was trained to improve book digitization by de-curling, a procedure which flattens images where pages are attached to the spine. The algorithm, trained on 1 million images, achieved over 90% accuracy in curling correction, and was 30 times faster than traditional workstations.
Start date: 01/11/2015
Duration in months: 18
Problem Description
Digitisation of books is crucial for commercialisation and preservation of older texts. The Bookscanner© product can automate the scanning process by physically turning pages. However, this process results in a 'curling' effect where the pages are attached to the spine.
Goals
New services
Challenges
The only current method for page curling correction is based on a projected laser grid that requires each page to be scanned twice. This is inefficient and often inaccurate. No AI-accelerated methods currently exist.
Innovation results
A Deep Neural Network was trained to de-curl 1 million images using simulated page curling. The algorithm uses cropped pages from a book image and creates an artificially curled page. The DNN can de-curl newly scanned pages with accuracy of over 90%. The training was 30 times faster using HPC.
Business impact
The CURLO solution, in collaboration with Arctur, can be offered as a post-processing service for Bookscanner© products, improving batch-mode scanning quality. This Software as a Service (SaaS) framework can also address digitisation needs in paperless economies like insurance and banking.