Tomato Fruit Detection and Counting in Greenhouses Using Deep Learning.

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Manual fruit counting is labor-intensive and limits phenotypic analysis.
• Automating fruit detection is crucial for efficient harvesting and crop management.
• Deep learning offers advantages over classical computer vision for variable agricultural data.

Purpose of the Study:

• To evaluate the MaskRCNN algorithm for tomato detection in greenhouse environments.
• To compare performance against existing methods using real-world image data.
• To assess the algorithm's capability for implicit depth learning for background elimination.

Main Methods:

• Utilized the MaskRCNN deep learning algorithm for object detection and pixel segmentation.
• Employed images captured in a greenhouse setting using a RealSense camera.
• Focused on detecting and counting tomatoes within the acquired images.

Main Results:

• Achieved detection and counting performance comparable to or exceeding prior studies.
• Demonstrated effectiveness even with lower resolution images compared to laboratory conditions.
• Confirmed MaskRCNN's ability to implicitly learn object depth for background separation.

Conclusions:

• MaskRCNN is a highly effective tool for automated tomato detection in greenhouses.
• The method shows promise for advancing precision agriculture and automated harvesting.
• Implicit depth learning capability enhances robustness in complex agricultural settings.