Distorted Visual Sequence Pattern Recognition

Read a 6-character code from noisy, distorted grayscale images using a CRNN + CTC deep-learning model. Scored by Character Error Rate (CER).

• Main deliverable: notebook/solution.ipynb — a self-contained, documented workflow that runs on Kaggle GPU, Colab, or locally.
• Beginner-friendly explanation + tuning guide: docs/DOCUMENTATION.md.
• Reusable code: the src/ package (same solution as clean modules).

Approach (TL;DR)

Image → CNN (visual features) → BiLSTM (sequence) → CTC (alignment-free text)

The CNN turns the image into a sequence of vertical "slices", the BiLSTM reads them in order, and CTC lets us train on the final string alone — no per-character boxes needed. See the docs for the full, plain-language walkthrough.

Dataset

• 20,000 train images + labels, 5,000 test images. Images are 200×100 grayscale.
• Labels are 6-char codes over a 31-symbol alphabet 23456789ABCDEFGHJKMNPQRSTUVWXYZ (ambiguous 0/1/I/L/O excluded).
• Expected layout:

data/cig_ps/
├── train_images/   # train-0.png ...
├── test_images/    # test-0.png ...
└── train-labels.csv

Download (one-off):

pip install gdown
python -c "import gdown; gdown.download(id='1e0Bpbmjp-Pc1Oz7luC1g8JhfzKpGZRHu', output='data/cig_ps.zip')"
python -c "import zipfile; zipfile.ZipFile('data/cig_ps.zip').extractall('data')"

Run on Kaggle (recommended — free GPU)

1. Create a Kaggle Dataset by uploading cig_ps.zip (it unzips to cig_ps/).
2. New Notebook → Add Data (your dataset) → Settings → Accelerator: GPU.
3. Upload notebook/solution.ipynb and Run All. The data path is auto-detected under /kaggle/input.
4. Download outputs/submission.csv, rename to submission_<name>_<enroll no.>.csv, and submit.

Run locally

pip install -r requirements.txt

# Option A: the notebook
jupyter notebook notebook/solution.ipynb

# Option B: the scripts
python -m src.train      # trains, saves outputs/crnn_best.pt
python -m src.predict    # writes outputs/submission.csv

CPU works but is slow (~17 min/epoch). Use a GPU (Kaggle/Colab) for real training.

Tuning

All hyper-parameters live in the notebook's CFG block (and src/config.py). The ranked, step-by-step tuning guide is in docs/DOCUMENTATION.md.