irfanbaloch3334/CNN_Segmentation_Model

🧾 Receipt (Bonnetje) Segmentation — YOLO11n-seg + Streamlit

A web app that detects receipts in photos and outlines them with a pixel-level segmentation mask, then lets you download the straightened (deskewed) receipt.

• Model: YOLO11n-seg, 1 class (bonnetje = receipt), ~2.8 M parameters, ~6 MB — ciCic/yolo11n-seg-bonnetjes on Hugging Face
• App: Streamlit · runs on CPU, no GPU needed

Features

• Upload one or many images (JPG, PNG, WEBP, BMP) or use the camera
• Adjustable confidence / IoU / image-size settings
• Annotated image with masks, plus a table (confidence, box, mask area)
• Extracts each receipt as a straightened crop (perspective-corrected), downloadable as PNG
• Handles phone-photo EXIF rotation and downsizes very large images automatically

Project structure

├── app.py                  # Streamlit UI
├── inference.py            # model loading, prediction, deskew helpers
├── requirements.txt        # Python dependencies (used by Streamlit Cloud)
├── packages.txt            # system libs for OpenCV on Streamlit Cloud
├── .streamlit/config.toml  # server / theme settings
├── models/                 # optional: put your own best.pt here
├── notebooks/              # original Colab training notebook
├── tests/                  # offline unit tests (pytest)
└── .github/workflows/ci.yml

Run locally

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt
streamlit run app.py

Open http://localhost:8501. The first start downloads the ~6 MB weights from Hugging Face and caches them.

Run the tests (no internet or GPU needed):

pip install -r requirements-dev.txt
pytest -q

Publish to GitHub

git init
git add .
git commit -m "Receipt segmentation Streamlit app"
git branch -M main
git remote add origin https://github.com/<your-username>/<repo-name>.git
git push -u origin main

Deploy on Streamlit Community Cloud (free)

1. Push the repo to GitHub (above). The repo must be public, or you must grant Streamlit access.
2. Go to https://share.streamlit.io → Create app → pick your repo and branch main.
3. Set Main file path to app.py.
4. Open Advanced settings → choose Python 3.12.
5. Click Deploy. The first build takes several minutes (PyTorch is large).

packages.txt installs the system libraries OpenCV needs; no other setup is required.

Using your own weights

Copy your trained checkpoint to models/best.pt (commit it; ~6 MB is fine for GitHub). It takes priority over the Hugging Face download. You can also set the environment variable MODEL_PATH, or HF_REPO_ID / HF_FILENAME to point at a different Hub model.

About the training notebook: the notebook fine-tunes on COCO128-seg, which has 80 classes, while this model has 1 (bonnetje). Ultralytics skips every label it can't map, so that run learns nothing (validation mAP ≈ 0). This app therefore uses the original Hugging Face model. To really fine-tune, use a receipt segmentation dataset with a single class 0: bonnetje and polygon labels.

Troubleshooting

Problem	Fix
ImportError: libGL.so.1	Make sure packages.txt is in the repo root (it is included).
Build fails / runs out of disk or memory on Streamlit Cloud	Default PyTorch from PyPI ships multi-GB CUDA libraries. Put these two lines at the top of requirements.txt to use the small CPU build: --extra-index-url https://download.pytorch.org/whl/cpu then torch==2.6.0+cpu and torchvision==0.21.0+cpu (each on its own line). Or deploy on Hugging Face Spaces / Render / Docker.
"Could not load the model"	First run needs internet access to Hugging Face, or provide models/best.pt.
No detections	Lower the confidence slider; make sure the receipt is clearly visible.

License

MIT for the app code (see LICENSE). Ultralytics is AGPL-3.0; the model weights follow the license on their Hugging Face page. Check both before commercial use.