2-Stage Neural Compression Pipeline

Noisy scanned document image → CNN OCR → Adaptive Huffman compression → lossless text recovery.

Two communicating Flask microservices built with PyTorch. Pre-trained weights are included. No training required to run the demo.

---

Repository Structure

pipeline/ ├── stage1_ocr/ ├── stage2_compression/ ├── launch.py ├── pipeline_demo.py ├── run_on_dataset.py ├── results.csv └── results_v2.csv

---

Dataset (for context)

This project uses: SimulatedNoisyOffice/

Used for:

• Training the denoiser
• Evaluating OCR under noisy conditions

---

Quick Start for Judges

Pre-trained weights are included.

1. Install dependencies

pip install -r requirements.txt

2. Run the entire system

python launch.py

That is all you need.

The script will:

• Start both microservices automatically
• Wait until they are ready
• Open a browser UI at http://localhost:5000

Press Ctrl+C to stop everything.

---

How to Use (UI)

1. Upload or drag an image
2. Optionally select a noise profile
3. Click Run Pipeline

---

What You Will See

Stage 1 OCR:

• Extracted text
• Character count
• Confidence
• Noise profile
• Denoiser usage

Stage 2 Compression:

• Compressed output
• Compression ratio
• Entropy
• Encoding efficiency

Final:

• Decompressed text
• Lossless match verification
• End-to-end latency

---

Architecture Overview

Stage 1 OCR:

• CNN: OCRNet (EMNIST-based)
• Denoiser: DnCNN
• Pipeline: image → denoise → segment → classify → text

Stage 2 Compression:

• Custom Adaptive Huffman Encoding (Vitter's Algorithm V)
• Fully lossless

---

Command-Line Demo (Optional)

This section is not required for judges.

Start services manually:

Terminal 1: cd pipeline/stage1_ocr python service.py --port 5001 --weights weights/ocr_model_best.pth --denoiser-weights weights/denoiser_best.pth

Terminal 2: cd pipeline/stage2_compression python service.py --port 5002

Then:

cd pipeline python pipeline_demo.py --images --noise none

---

Full Dataset Evaluation

cd pipeline python run_on_dataset.py --csv results.csv

Evaluates all images in SimulatedNoisyOffice and writes metrics.

---

Notes for Judges

• One command to run everything
• No training required
• Includes OCR + denoising + compression
• Custom Huffman implementation
• Lossless recovery verified

Built With

• cnn
• dncnn
• flask
• numpy
• pillow
• python
• pytorch