Total records
111,143
Distributed Urdu OCR
Restore and digitize degraded Urdu records at national scale.
A deep-learning pipeline combines SMP U-Net restoration with Conv-Transformer OCR, then executes inference over distributed Hadoop workers to process large document batches in parallel.
تاریخی اور سرکاری اردو دستاویزات کی خودکار بحالی اور ڈیجیٹل تبدیلی
Final CER
8.14%
Final WER
18.33%
Restoration quality
PSNR 34.52 dB / SSIM 0.9932
Section 2
Data + Preprocessing
111,143 records combine printed Nastaleeq and handwritten Urdu. Inputs are standardized to 1 x 128 x 2048 grayscale tensors before restoration and OCR.
Dataset composition
MMU-OCR-21 (Printed Nastaleeq)100,541
NUST-UHWR (Handwritten Urdu)10,602
Train / Val / Test split
Train
77,800
70%
Validation
16,671
15%
Test
16,672
15%
Standardization pipeline
- 1Grayscale conversion
- 2Resize to fixed height: 128
- 3Aspect-ratio-preserving right padding to width: 2048
Synthetic degradations used
Applied transformations used to stress-test restoration robustness
Gaussian BlurGaussian NoiseSalt & PepperLow ContrastAffine SkewPadding Crop
Section 3
Deep Learning Pipeline
Noisy Input
Degraded scanned page
SMP U-Net Restoration
ResNet34 encoder
Restored Image
PSNR 34.52 / SSIM 0.9932
Conv-Transformer OCR
Urdu sequence decoding
Model block
Restoration block
SMP U-Net + ResNet34 encoder
Params 24,430,097MSE 0.000574PSNR 34.52 dBSSIM 0.9932
- ImageNet-pretrained encoder for stable low-level feature extraction.
- Restores contrast and edge fidelity before OCR decoding.
- Optimized for degraded archive and government document scans.
Model block
OCR block
Conv-Transformer sequence recognition
Params 7,337,197Vocab 1737 CNN blocks3 Encoder / 3 DecoderCross-Entropy objective
- CNN backbone encodes visual tokens from restored Urdu text lines.
- d_model=256, nhead=8, feedforward=1024
- Transformer decoder outputs character sequence without CTC.
Section 4
Results + Benchmarking
Final OCR test metrics
Loss
0.2075
CER
8.14%
WER
18.33%
DL vs Tesseract benchmark
CER improvement: 58.29 ppWER improvement: 82.83 pp
- Tesseract: CER 62.04%, WER 93.15%
- DL model: CER 3.75%, WER 10.32%
Error-rate comparison
Handwriting challenge note
UHWR handwritten mini-sample remains harder: CER 29.89% and WER 55.29%, reinforcing handwriting as the current difficult regime.
Section 5
Distributed Inference Architecture
Execution flow
Uploaded batches enter HDFS, MapReduce shards page groups across workers, each worker runs the same preprocessing → restoration → OCR block, and reducer stages merge outputs into ordered Urdu text files.
Input batch→Mapper nodes→Parallel inference→Reducer→Output bundle
Tech stack
React/Next.js
FastAPI
PyTorch
Hadoop/HDFS/MapReduce
Docker
Runport
Graph composition from distributed architecture
Press enter or space to select a node. You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.
Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.
Scroll to explore
Interactive demo simulator
Upload → Process → Review Urdu OCR Output
Input panel
Upload images
Line segmentation
Processing panel
Distributed scheduler (2 Hadoop data nodes)
Upload a zip to start the lifecycle: idle → uploading → processing → completed | failed.
Results panel
Output mapping and Urdu text viewer
Run a job to view 1:1 output mapping (A.pdf → A.txt), then inspect line-level OCR text.