Single-Shot Convolution Neural Networks for Real-Time Fruit Detection Within the Tree.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Traditional fruit detection methods using hard-coded features are accurate but computationally intensive and slow for real-time applications.
• Deep learning offers a way to automate feature extraction, eliminating the need for manual feature engineering for diverse fruit attributes.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for accurate and efficient fruit detection and counting in trees.
• To enable real-time fruit detection for applications such as automated harvesting.

Main Methods:

• Utilized a deep convolutional neural network architecture based on single-stage detectors.
• Implemented an image processing technique that divides images into a grid for localized detection and localization.
• Trained the network on over 5000 images of apples and pears, with manually labeled bounding boxes.

Main Results:

• Achieved over 90% accuracy in fruit detection (apples and pears).
• The system processes images at a speed exceeding 20 frames per second (FPS), suitable for real-time applications.
• Developed a model to account for detection errors based on visible vs. actual fruit counts.

Conclusions:

• Deep learning, specifically single-shot detectors, provides an effective solution for fruit detection and counting within tree canopies.
• The developed system offers a significant speed improvement over traditional methods, making it viable for robotic applications.