Real-Time Grading of Defect Apples Using Semantic Segmentation Combination with a Pruned YOLO V4 Network

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Traditional apple grading systems using belts or rollers cause bruising, leading to economic losses.
• High-quality apples require gentle handling and accurate defect detection for efficient grading.

Purpose of the Study:

• To develop a real-time apple detection and classification system that minimizes fruit damage.
• To improve the accuracy of identifying and quantifying defects in apples.

Main Methods:

• Designed separate fruit trays to prevent bruising during image acquisition.
• Employed BiSeNet V2 for semantic segmentation of defective apple areas, achieving 99.66% MPA.
• Utilized model pruning on the YOLO V4 network to enhance defect detection accuracy.
• Developed a surface mapping method for precise defect area calculation.

Main Results:

• The BiSeNet V2 model outperformed DAnet and Unet in defect segmentation.
• The pruned YOLO V4 network further improved defect region detection accuracy.
• The system achieved an average apple classification accuracy of 92.42% with an F1 score of 94.31%.

Conclusions:

• The proposed system effectively reduces apple bruising and enhances grading accuracy.
• The developed deep learning models and surface mapping method show significant potential for commercial apple grading machines.