YOLO-Tomato: A Robust Algorithm for Tomato Detection Based on YOLOv3.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Automatic fruit detection is crucial for harvesting robots but faces challenges from environmental factors like illumination variations, occlusion by branches/leaves, and fruit overlap.
• Existing fruit detection models struggle with the complexity of real-world agricultural environments.

Purpose of the Study:

• To develop an improved tomato detection model, YOLO-Tomato, based on YOLOv3, to address the challenges in automatic fruit detection for harvesting robots.
• To enhance the accuracy and efficiency of fruit localization in complex agricultural settings.

Main Methods:

• An improved tomato detection model, YOLO-Tomato, was developed based on the YOLOv3 architecture.
• A dense architecture was incorporated into YOLOv3 to improve feature reuse and model compactness.
• Traditional rectangular bounding boxes (R-Bbox) were replaced with circular bounding boxes (C-Bbox) for more precise tomato localization and improved Intersection-over-Union (IoU) calculations during Non-Maximum Suppression (NMS).

Main Results:

• The YOLO-Tomato model demonstrated improved detection performance compared to several state-of-the-art methods.
• The modifications, including the dense architecture and circular bounding boxes, were validated through an ablation study.
• The use of C-Bbox improved IoU calculation and reduced prediction coordinates, leading to more accurate localization.

Conclusions:

• The proposed YOLO-Tomato model significantly enhances automatic tomato detection capabilities in challenging agricultural environments.
• The integration of a dense architecture and circular bounding boxes offers a promising approach for robust fruit detection in robotic harvesting systems.
• YOLO-Tomato achieved superior detection performance, making it a valuable tool for advancing agricultural automation.