Apple Detection using YOLO and Faster R-CNN

Overview

This project implements an apple detection system using two different object detection models: YOLO (You Only Look Once) and Faster R-CNN. The models are trained on a custom dataset and used for detecting apples in images and video streams. The system is designed to process images, apply non-maximum suppression (NMS) for filtering detections, and save annotated results.

Features

• Apple detection using YOLOv8 and Faster R-CNN
• Training with a custom dataset
• Evaluation metrics: [email protected], [email protected]:0.95, Precision, Recall
• Non-Maximum Suppression (NMS) for filtering overlapping detections
• Batch processing of test images
• Saves annotated images with detection results

Dataset

The dataset is defined in apple.yaml, which contains the paths to training and validation image datasets. The dataset includes:

• Bounding box annotations for apples in each image.
• RGB images collected from orchards.

Dependencies

Ensure you have the following dependencies installed:

pip install ultralytics opencv-python numpy torch torchvision

Model Training

Train YOLOv8 Model

from ultralytics import YOLO

model = YOLO("/path/to/yolov8s.pt")  # Pretrained model
model.train(data="/path/to/apple.yaml", epochs=100, batch=32, imgsz=640)

Train Faster R-CNN Model

checkout https://github.com/nicolaihaeni/MinneApple/tree/master for more information on training FRCNN and acquire data.

Object Detection and Annotation

The detection pipeline applies NMS to filter redundant detections and then annotates the detected apples in images.

Example Results

Future Improvements

• Fine-tuning YOLO and Faster R-CNN with additional orchard datasets
• Implementing real-time apple detection on video feeds
• Deploying on Jetson Nano/AGX Orin for edge computing

Author

This project is developed for advanced apple detection applications in agricultural automation.