Arsalan Younus.
Back to Projects

Document Denoising Pipeline

Deep learning preprocessing that cleans noisy scans while preserving text clarity, significantly improving downstream OCR accuracy on poor-quality documents.

The Business Problem

Poor-quality scans (shadows, blur, low contrast) significantly hurt OCR accuracy. Generic image filters degraded text or did not address the right type of degradation.

The client needed a dedicated preprocessing step that cleans noisy scans without losing text clarity.

The Technical Solution

I built a U-Net-based denoising pipeline that removes shadows, corrects lighting, reduces blur, and enhances text clarity, tuned to preserve text while suppressing noise.

The system runs as a preprocessing step before OCR, integrated with the noise classification stage for intelligent routing.

The Scalability Factor

Runs as a GPU-accelerated preprocessing service in the document intelligence pipeline. Tuned for production volume with balanced GPU usage and latency.

Business Impact

19% OCR accuracy improvement on noisy pages; plug-in preprocessing stage in production.

Fewer documents rejected for low quality; manual rescan requests decreased.

Built with

U-Net
Deep Learning
PyTorch
OpenCV
Image Processing