🌱 Plant Counting Framework An end-to-end pipeline for counting plants in drone imagery — from raw aerial images through automated filtering, manual labeling, and COCO/YOLO annotation export.

Overview This framework was built to process high-resolution DJI drone photographs of crop fields. It automates the tedious parts of dataset creation: filtering out irrelevant image tiles using HSV-based green detection, presenting candidate patches to a human labeler via a web UI, storing labels in MongoDB, and exporting the result as YOLO-format annotations ready for model training.

Pipeline DJI Drone Images

  │

  ▼

1. Green Coverage Analysis (start.py)

   • Convert to HSV color space

   • Detect green pixels (plants)

   • Filter images with < 20% green coverage

   • Plot coverage distribution (mean / std)

     │

     ▼

2. Image Tiling & Filtering (webapp.py)

   • Split full-res DJI images into 60×80 px patches

   • Drop tiles with < 20% green coverage

     │

     ▼

3. Web-based Labeling UI (webapp.py + Flask)

   • Serve each patch to a human annotator

   • Label: plant present (yes) or not (no)

   • Store bounding box + label in MongoDB Atlas

     │

     ▼

4. Annotation Export (/convert endpoint)

   • Read labels from MongoDB

   • Write YOLO-format .txt annotation files

   • Copy images into coco/images/ directory

     │

     ▼

COCO/YOLO Dataset — ready for model training

Features HSV green detection — automatically discards image tiles that contain no meaningful plant coverage Tile splitting — chops large drone frames into 60×80 px patches for patch-level labeling Flask labeling app — browser-based UI to label patches one at a time; no annotation software required MongoDB Atlas storage — labels and bounding boxes are persisted to a cloud database YOLO export — converts stored labels to normalized YOLO .txt format alongside copied images Statistics & visualization — start.py plots the green-coverage distribution across a dataset

Tech Stack Component Technology Image processing OpenCV, Pillow Web framework Flask Database MongoDB Atlas (flask_pymongo) Visualization Matplotlib, SciPy Image source DJI drone (field/aerial) Label format YOLO (normalized bounding boxes)

Project Structure Plant-counting-Framework/

├── images/ # Raw DJI drone images (input)

├── coco/

│ ├── images/ # Copied images for training

│ └── data/ # YOLO annotation .txt files

├── serving_static/

│ └── static/

│ ├── images/ # Tiled patches served to labeler

│ └── script/ # Frontend JavaScript

├── templates/ # Flask HTML templates

├── Gmaps/ # Google Maps integration

├── GoogleEarth/ # Google Earth imagery

├── start.py # Green coverage analysis & filtering

├── webapp.py # Flask labeling app + export

├── convertCoco.py # COCO conversion utility

├── readProp.py # Property reader

└── split.py # Image splitting utility

Getting Started

1. Clone the repository git clone https://github.com/kvsandy/Plant-counting-Framework.git

cd Plant-counting-Framework 2. Install dependencies pip install flask flask-pymongo opencv-python pillow scipy matplotlib numpy 3. Add your drone images Place your DJI images inside the images/ directory:

images/

├── DJI_0444.JPG

├── DJI_0446.JPG

└── ... 4. Analyze green coverage (optional pre-filter) python start.py

This scans the images2/ folder, removes images with less than 20% green coverage, and plots the distribution of the remaining images. 5. Configure MongoDB Update the connection string in webapp.py to point to your own MongoDB Atlas cluster:

app.config["MONGO_URI"] = "mongodb+srv://:@.mongodb.net/label_db" 6. Run the labeling web app python webapp.py

Open http://localhost:5000 in your browser. The app will:

Automatically tile your DJI images into 60×80 px patches Filter out tiles with insufficient green coverage Present each patch for you to label as plant or not plant 7. Export to YOLO format Once labeling is complete, visit:

http://localhost:5000/convert

This reads your labels from MongoDB and writes YOLO-format annotation files to coco/labels/ alongside the source images in coco/images/.

Label Format Exported annotations follow the YOLO format:

<class_id> <x_center> <y_center>

0 = plant present 1 = no plant All values are normalized to [0, 1] relative to the source image dimensions.

Requirements Python 3.7+ MongoDB Atlas account (free tier works) DJI drone images (or any high-resolution field photos)

Contributing Pull requests are welcome. For major changes, please open an issue first to discuss what you'd like to change.

License MIT