Smart OCR - Computer Vision & NLP
Many institutions parse and store tons of forms and other form of official documentation. All these forms have to be analyzed and stored in a meticulous fashion and sometimes due to the variations of hand-writings, processing and cataloguing them is a very onerous task.
Optical Recognition Systems can be used to extract the data but the models used currently are serviceable at best. Our model, utilizing modern developments in image processing and Natural Language Processing sets out to create a more resolute model.
The Philippines
Client
Content and Data Management Company
CLIENT
Our client is a content and data management company based in the Philippines. Their roots are based in Ed-Tech services providing measures and methodologies for many international conglomerates involved with this sector. We were tasked by our clients to create an AI-affiliated OCR module for extracting data from an array of forms provided by them. These forms contain various amounts of data, languages, and scripts.
The extraction of these data is vital for the functioning of the company since they contain important financial and personal data. With the recent developments in AI technologies and our expertise in this field, we can create a solid solution that provides the company with vast support functions.
CHALLENGE
A massive array of challenges regarding this model that need to be tackled are listed below :
-
Documentation Backgrounds - Some forms, documents, and identity cards can have varying colored backgrounds. These can be a major hindrance to the OCR model facilitating erroneous character identification.
-
Cursive Handwriting & Unconventional Calligraphy Extraction - Cursive and unconventional calligraphy has been a major problem for OCR for ages. Their uncommon aspect can create major disparity, and could sometimes need manual intervention for resolution.
-
Scan Quality - The quality of scans can sometimes heavily affect the OCR process. It has to be ensured that scanned image has to be without blemishes and has to be at a decent resolution.
SOLUTION
We have arrived at a solution with the project basis and our own research regarding the challenges in mind. They are listed below:
-
Using pre-configured datasets, we can arrive at a method to identify the English-Latin Script characters from handwriting.
-
The derived dataset is used as a base to construct a neural network model, which can do comparative analysis between the input and our methodology.
-
The image processing operation contains various logical steps so that we could arrive at more clear data. Image Binarization (conversion of a multicolored imaged into a duo-tone image containing only black and white), denoising (removal of visual noise and other artifacts from the scanned image), and various other operations are run to arrive at cohesive data.
-
Natural Language Processing is done to find various strategies for auto-completion. Identities, spellings (also alternative ones) and addresses are extracted for this purpose. This can further exact and context-specific identification.
RESULT
With the arrival of model like this, which contains not only the aspects of traditional OCR methods but also has the support of integrated AI-oriented image and language analysis, the problems of redundancy, illegibility, and obscurity are solved with greater ease. Data storage and retrieval is also made easy because of digitization and indexing done by the OCR extraction operations. Compartmentalization of data is also fine-tuned to fit the rules of the Client. Efficiency is boosted and the need for manual rechecking is reduced drastically.