TomatoDet: Anchor-free detector for tomato detection
Guoxu Liu1, Zengtian Hou2, Hongtao Liu1
1Goertek College of Science and Technology Industry, Weifang University, Weifang, China.
Frontiers in Plant Science
|August 22, 2022
Summary
A new detector, TomatoDet, enhances automatic greenhouse harvesting by accurately identifying tomatoes despite challenging conditions like poor lighting and occlusion. This robust fruit detection system promises to advance commercial robotic harvesting applications.
Area of Science:
- Agricultural Engineering
- Computer Vision
- Robotics
Background:
- Accurate fruit detection is crucial for automated harvesting robots in greenhouses.
- Environmental challenges like uneven illumination, occlusion, and fruit overlap impede robust detection systems and commercialization.
Purpose of the Study:
- To develop an improved anchor-free fruit detection system, named TomatoDet, to address challenges in greenhouse environments.
- To enhance feature expression and simplify regression for more accurate tomato detection.
Main Methods:
- Incorporated an attention mechanism into the CenterNet backbone for improved feature representation.
- Introduced a novel circle representation to optimize the detector for tomato shape and simplify keypoint regression.
Main Results:
- TomatoDet outperformed state-of-the-art detectors in tomato detection, achieving an F1 score of 95.03% and an average precision of 98.16%.
- The detector demonstrated robust performance under varying illumination and occlusion conditions.
Conclusions:
- TomatoDet offers a promising solution for robust and accurate tomato detection in complex greenhouse environments.
- The proposed method facilitates the commercial application of automated harvesting robots.
