Appearance quality classification method of Huangguan pear under complex background based on instance segmentation...

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning applications in agriculture

Background:

• Automated grading of 'Huangguan' pear appearance quality is crucial for fruit processing.
• Traditional computer vision methods face limitations in detecting diverse pear shapes and disease spot types.
• Deep learning, particularly convolutional neural networks, offers a promising solution for accurate disease detection.

Purpose of the Study:

• To develop an integrated deep learning framework for automated 'Huangguan' pear disease spot detection and grading in complex backgrounds.
• To evaluate the performance of different deep learning models for instance segmentation, semantic segmentation, and classification stages.

Main Methods:

• An integrated framework combining instance segmentation (Mask R-CNN), semantic segmentation (DeepLabV3+, UNet, PSPNet), and grading models (ResNet50, VGG16, MobileNetV3).
• Mask R-CNN with CLAHE preprocessing for initial pear segmentation from complex backgrounds.
• DeepLabV3+, UNet, and PSPNet for segmenting disease spots and calculating spot-to-pear area ratios for grading.

Main Results:

• The Mask R-CNN with CLAHE preprocessing achieved 97.38% pixel accuracy (PA) and 68.08% Dice coefficient for pear segmentation.
• DeepLabV3+ demonstrated the highest accuracy in spot segmentation with 94.03% PA and 67.25% Dice coefficient.
• ResNet50 excelled in grading, achieving 97.41% average precision (AP) and 95.43% F1 score.

Conclusions:

• The proposed framework effectively segments 'Huangguan' pears and their disease spots in complex environments.
• The integrated deep learning approach enables accurate grading of fruit disease severity.
• This study provides a novel framework and theoretical foundation for automated pear disease assessment and grading.