Pest_Detection

This project aims to implement crop pest object detection based on Faster R-CNN (with ResNet-50-FPN as the backbone network) using PyTorch. The project includes a baseline model and an improved version, the latter of which adopts a more refined training strategy, focusing on enhancing the accuracy of category recognition.

1. Project structure

Pest_Detection/
└── CNN/
    ├── Faster_R-CNN/              
    │   ├── runs/                
    │   ├── Faster_R-CNN.py    
    │   ├── test.py               
    │   └── train.py             
    ├── Faster_R-CNN_Improved/   
    │   ├── runs/                
    │   ├── best_faster_rcnn_weights.pt
    │   ├── Faster_R-CNN_Improved.py
    │   ├── test.py               
    │   └── train.py               
    ├── output/                  
    ├── data.yaml                  
    └── detect.py

2. Environmental Requirements and Dependencies

This project is based on the PyTorch deep learning framework and uses torchmetrics and sklearn for evaluation, as well as albumentations for data augmentation.

pip install torch torchvision torchaudio 
pip install numpy opencv-python Pillow
pip install torchmetrics scikit-learn albumentations matplotlib tensorboard

Category	Library	Purpose	Installation Command (suggested)
Deep Learning	torch	Core PyTorch framework.	pip install torch torchvision torchaudio
	torchvision	Contains the Faster R-CNN model, pre-trained weights, and image transformations.	(Included with torch)
Data Handling	numpy	Fundamental package for numerical operations and array handling.	pip install numpy
	opencv-python	Used by the AgroPestYOLODataset class for image loading and BGR to RGB conversion.	pip install opencv-python
Augmentation	albumentations	Used for efficient and flexible data augmentation pipelines in the AgroPestYOLODataset.	pip install albumentations
Evaluation	torchmetrics	Used to calculate mAP metrics during validation and testing.	pip install torchmetrics
	scikit-learn	Used to calculate standard classification metrics (Precision, Recall, F1, Accuracy) in the evaluate_model function.	pip install scikit-learn
Logging/Vis	tensorboard	Used by train.py for tracking training loss and validation metrics.	pip install tensorboard
	matplotlib	Likely required by detect.py for displaying inference results.	pip install matplotlib
Other	Pillow (PIL)	Common dependency for image processing libraries.	pip install Pillow

3. Model Training

3.1 Baseline Model (Faster_R-CNN.py)

• Configuration: Strategy
  • Backbone: ResNet-50-FPN
  • Optimizer: SGD (LR: 0.005, Momentum: 0.9, Weight Decay: $5 \times 10^{-4}$)
  • LR Scheduler: StepLR (LR drops by a factor of 10 every 3 epochs)
  • Data Augmentation: Includes geometric transformations (flip, rotation, scale) and pixel-level transformations

3.2 Improved Model (Faster_R-CNN_Improved.py)

• Configuration: Strategy
  • Backbone: ResNet-50-FPN
  • Optimizer: SGD (LR: 0.001, Momentum: 0.9, Weight Decay: $5 \times 10^{-4}$)
  • LR Scheduler: CosineAnnealingLR (Min LR: $1 \times 10^{-6}$)
  • Data Augmentation: Only pixel-level transformations retained (removed geometric to stabilize bounding box regression)
  • Gradient Clipping: max_norm = 5.0

Running Faster_R-CNN Training

The training script saves the model weights with the highest mAP@0.5 on the validation set to best_faster_rcnn_weights.pth.

# Run training for the basic Model (Logs indicate you ran this version)
python Faster_R-CNN\train.py

Running Faster_R-CNN Testing

# Run testing for the basic Model (Logs indicate you ran this version)
python Faster_R-CNN\test.py

Running Faster_R-CNN_Improved Training

The training script saves the model weights with the highest mAP@0.5 on the validation set to best_faster_rcnn_weights.pth.

# Run training for the Improved Model (Logs indicate you ran this version)
python Faster_R-CNN_Improved\train.py

Running Faster_R-CNN_Improved Testing

# Run testing for the Improved Model (Logs indicate you ran this version)
python Faster_R-CNN_Improved\test.py

4. Improved Model Training Results Summary (30 Epochs)

Metric	Best Validation Performance (Epoch 13)	Final Test Set Performance
Detection mAP@0.5	0.7352	0.7476
mAP (COCO)	N/A	0.3987
F1 Score (TP-based)	0.7920	0.7713
Total Training Time	N/A	13h 30m 35s

5. Model Inference and Visualization (detect.py)

detect.py is used to load the trained weights and perform detection and visualization on a single image.

Running Inference Verify Weights: Ensure best_faster_rcnn_weights.pth is present in the project root directory.

Update Class Names: Make sure CLASS_NAMES in detect.py is updated as required in Section 3.

Specify Image Path: Modify test_image_path at the bottom of the detect.py script.

# Example: Run inference and display results
python detect.py

The script will load the model, perform inference using the predefined threshold (default is 0.5), and display the result in a Matplotlib window.