Application of deep learning for fruit defect recognition in Psidium guajava L

Area of Science:

• Agricultural Science
• Computer Vision
• Food Quality Control

Background:

• Psidium guajava L. (guava) is a key tropical and subtropical fruit, with year-round production in Taiwan necessitating quality standardization.
• Appearance defects significantly impact the marketability and standardization of harvested guavas.

Purpose of the Study:

• To develop and evaluate an automated system for detecting and classifying appearance defects on guavas.
• To improve the efficiency and accuracy of quality control in guava harvesting and grading.

Main Methods:

• Utilized the YOLO v4 pretrained network architecture for defect detection.
• Collected 1701 images of 189 Psidium guajava L. fruits from various farms, categorizing defects into thirteen classes.
• Assessed model performance using metrics such as false positive rate, false negative rate, and accuracy.

Main Results:

• The YOLO v4 model achieved an overall accuracy of 88.15% in defect detection.
• Demonstrated low false positive (6.62%) and false negative (5.03%) rates for general defects.
• Achieved less than 9% false positive and false negative rates for specific fungal diseases (Colletotrichum gloeosporoides, Pestalotiopsis psidii, Phyllosticta psidiicola).
• Showcased real-time applicability with a minimum detectable defect size of 13×14 pixels and a processing speed of 12 FPS.

Conclusions:

• The YOLO v4 model is highly effective for automated detection of appearance defects in Psidium guajava L.
• The system's performance supports its implementation in real-time harvesting and grading processes for enhanced quality control.
• This AI-driven approach offers a robust solution for standardizing guava quality.