A fluorescence detection method for postharvest tomato epidermal defects based on improved YOLOv5m.

Area of Science:

• Agricultural Science
• Computer Vision
• Food Science

Background:

• Visual grading of tomatoes is challenged by smooth skin, uneven lighting, and hidden defects.
• Intelligent detection of postharvest tomato epidermal defects can enhance economic value.

Purpose of the Study:

• To develop an intelligent system for detecting postharvest tomato epidermal defects.
• To improve the accuracy and efficiency of tomato quality grading.

Main Methods:

• Designed a fluorescence imaging device for capturing tomato skin defect datasets.
• Improved the YOLOv5m model by integrating Convolutional Block Attention Module and Switchable Atrous Convolution.
• Evaluated the model's performance through comparison and ablation experiments.

Main Results:

• The improved YOLOv5m model achieved 89.93% Precision, 82.33% Recall, and 87.57% mean Average Precision.
• Outperformed YOLOv5m, Faster R-CNN, and YOLOv7 in detection accuracy.
• Reduced average detection time by 47.04 ms per picture.

Conclusions:

• The fluorescence imaging and improved YOLOv5m model effectively detect various tomato epidermal defects, including imperceptible ones.
• Provides robust technical support for intelligent tomato detection and quality grading.