Automated Detection of Nematode Eggs

This repository contains the main scripts and documentation for an automated deep learning system to detect nematode eggs in microscopic images. The project was developed in collaboration with Lincoln University (Christchurch, NZ), in partnership with the Department of Agricultural Sciences, which focused on improving livestock health research, the goal was to streamline the egg detection process through a computer vision-based solution.

The project supports training, fine-tuning, evaluation, prediction, and visualisation for Faster R-CNN, YOLOv8s and DeepLabV3+ models.

---

Folder Structure

This Git repository does not include image, CSV, or other data files. To access these, please send a request to shionshine@gmail.com.

automated-detection-of-nematode-Eggs-/
├── Data/                  # Raw microscopic images and annotation files
├── dataset/               # Processed datasets (train/val/test splits, YOLO labels, DeepLab masks etc.)
│   ├── test
│   │   ├── annotaions     # .xml files for bounding box annotations
│   │   ├── images         # .tif image files
│   │   ├── json           # .json files. Faster RCNN training requires.
│   │   ├── labels         # .txt files. YOLO training requires.
│   │   └── masks          # .png files. DeepLab training requires. 
│   ├── train
│   │   └── ...
│   ├── val
│   │   └── ...
├── DeepLab/               # DeepLabV3+ segmentation scripts and configs
│   ├── deeplab_training.py        # Training or fine-tuning 
│   ├── evaluate_visual_deeplab.py # Perform evaluation + visualisation 
│   ├── inference_deeplab_model.py # Run prediction separately if needed (it's usually integrated with training)
│   └── ...
├── Faster_rcnn/           # Faster R-CNN detection scripts and configs
│   ├── faster_rcnn.py
│   ├── evaluate_visual_faster_cnn.py
│   ├── inference_faster_rcnn.py
│   └── ...
├── YOLO/                  # YOLOv8 training scripts
│   ├── yolo_training.py
│   ├── evaluate_YOLO.py
│   ├── visual_yolo_prediction.py 
│   └── ...
├── Processed_Images/      # Model predictions and visualisations
├── model/                 # Saved model weights and checkpoints
├── Helper/                # Utility functions, particularly for converting .tif image files into formats suitable for training
│   ├── split_prep_sample.py
│   └── ...
├── evaluation/            # Evaluation metrics, running logs and Tensorboard logs
├── README.md              # Project documentation (this file)
└── requirements.txt       # Python dependencies

---

Main Components

1. **Data Preparation for Training **

• Script in Helper/ to split raw data into train/val/test sets
• Convert images to formats that suitable for training (e.g., Pascal VOC XML to YOLO readable .txt, etc.).

2. Model Training & Tuning

• Faster R-CNN:
  • Faster_rcnn/faster_rcnn.py for training and fine-tuning with different learning rates and backbones.
• YOLOv8:
  • YOLO/yolo_training.py for training YOLOv8 models with various hyperparameters and optimisers.
• DeepLabV3+:
  • DeepLab/deeplab_training.py for segmentation training.

3. Inference & Prediction

• inference_faster_rcnn.py for example, is a script for running inference on trained models and saving predictions. It is built to run inference separately if needed, which usually is already integrated into the training pipeline.
• Outputs are typically saved in /Processed_Images

4. Evaluation & Visualization

• DeepLab/evaluate_visual_deeplab.py, faster_rcnn/evaluate_visual_faster_cnn.py etc for evaluating predictions (precision, recall, F1, mAP, PR curves). Outputs are typically saved in evaluation.
• For YOLO, TensorBoard was used to compare different YOLO model variants. See more https://www.tensorflow.org/tensorboard/get_started

---

How to Use

Install Dependencies

Ensure Python 3.11.9 is installed, then install the required packages:

   pip install -r requirements.txt

---

Debug notes

• Adjust paths and hyperparameters as needed for your experiments.
• Refers to Helper/label_instruction.md if you have any questions around labeling.

---

Contact

For questions or contributions, please contact Jacquelin Ruan at shionshine@gmail.com or open an issue or pull request.

---