Records Digitization with Computer Vision

Problem

Manual paper records create bottlenecks: they are hard to search, easy to lose, and expensive to aggregate for reporting. Organisations need a reliable path from physical documents to structured electronic records.

Context

A scholarly project automating the conversion of manual data records into electronic records using convolutional neural network-based computer vision, deployed as both an Android application and a web-based platform.

Dataset

Scanned and photographed manual record forms across varied handwriting, layout and capture conditions typical of institutional record-keeping workflows.

Methodology

End-to-end pipeline from document capture through CNN-based field detection, classification and extraction into structured electronic records, with separate mobile and web interfaces for different operational contexts.

Model / architecture

Convolutional neural networks for document layout understanding, field detection and character or token classification, optimised for deployment on mobile devices and web servers.

Results

Operational Android and web applications that reduce manual transcription, improve record retrieval and create a structured data layer for downstream analytics and reporting.

Evaluation metrics

Modality

Document imagery

Architecture

CNN-based extraction

Deployment

Android app and web application

Challenges

Handwriting variability, inconsistent form layouts, image quality from phone cameras, and the need for human-in-the-loop verification on high-stakes fields.

Impact

Faster digitisation of institutional records, reducing administrative burden and unlocking analytics that depend on structured electronic data.

Tools / technology

• Python
• TensorFlow
• CNN
• Android
• Django
• Flask