Toward Real Scenery: A Lightweight Tomato Growth Inspection Algorithm for Leaf Disease Detection and Fruit Counting.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Monitoring tomato plant growth faces challenges from disease variability and environmental conditions.
• Intelligent surveillance systems require robust methods for simultaneous disease identification and fruit counting.

Purpose of the Study:

• To develop an integrated cascade framework for autonomous tomato crop surveillance.
• To enhance deep learning models for accurate leaf disease detection and fruit counting.

Main Methods:

• An autonomous robot with a smartphone camera collected greenhouse images.
• The YOLO-TGI deep learning network was improved with Ghost and CBAM modules.
• State-of-the-art trackers (Byte-Track, Motpy, FairMot) were integrated for fruit counting.

Main Results:

• The YOLO-TGI and Byte-Track combination showed the best performance.
• YOLO-TGI-N model achieved low computational demands (2.05 G FLOPs, 3.7 M weights) with a 0.72 mAP for disease detection.
• YOLO-TGI-S and Byte-Track achieved high accuracy (R²=0.93, RMSE=9.17) for fruit counting with enhanced speed.

Conclusions:

• The developed framework offers a potential solution for automated surveillance in agriculture.
• This approach can be adapted for monitoring a wide range of fruit and vegetable crops.