FAR-Net: Feature-Wise Attention-Based Relation Network for Multilabel Jujube Defect Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Production processes can result in products with multiple defects.
• Accurate identification of all defects is crucial for improving agricultural planting and production.
• Standard convolutional neural networks may overlook semantic correlations between multiple defect labels.

Purpose of the Study:

• To propose a novel network architecture for multilabel defect classification.
• To address the limitations of existing methods in capturing label correlations.
• To improve the accuracy of identifying multiple surface defects on jujubes.

Main Methods:

• Developed a Feature-wise Attention-based Relation Network (FAR-Net).
• FAR-Net incorporates modules for image feature extraction, label-wise aggregation, feature activation/deactivation, and correlation learning.
• A unique multilabel jujube defect dataset was created for evaluation.

Main Results:

• The proposed FAR-Net demonstrated improved performance over the Inception v3 backbone.
• Average precision for three main composite defects increased by 5.77%, 4.07%, and 3.50%.
• The relation learning mechanism effectively captured correlations between defect labels.

Conclusions:

• FAR-Net significantly enhances multilabel classification accuracy for jujube defects.
• The network's ability to learn label correlations is key to its improved performance.
• This method offers practical value for improving agricultural production processes through better defect identification.