Related Experiment Video
Updated: May 31, 2026
08:27
Image Recognition and Parameter Analysis of Concrete Vibration State Based on Support Vector Machine
Published on: January 5, 2024
Class prototype rectification and multi-scale feature measurement for few-shot classification of bearing surface
Yan Cang1, Chunguang Li2, Xuanshang Zhang3
1College of Information and Communication, Harbin Engineering University, Harbin, 150000, China. cangyan@hrbeu.edu.cn.
Scientific Reports
|May 28, 2026
Summary
This study introduces a meta-learning framework to enhance few-shot industrial defect detection accuracy with limited data. The approach improves classification by learning adaptive embeddings and robust prototypes, crucial for manufacturing quality control.
Area of Science:
- Computer Vision
- Machine Learning
- Industrial Engineering
Background:
- Industrial defect detection struggles with limited generalization and accuracy when labeled data is scarce.
- Few-shot learning presents a significant challenge in scenarios with insufficient annotated samples for training robust models.
Purpose of the Study:
- To develop a novel meta-learning framework for few-shot industrial defect detection.
- To enhance classification performance and robustness in data-scarce industrial environments.
Main Methods:
- Proposes a meta-learning framework incorporating task-adaptive embeddings, rectified class prototypes, and multi-scale metric inference.
- Utilizes adaptive weighted pooling, local sliding operations, Class Prototype Rectification (PR) with Sum-of-Absolute-Differences (SAD), and a Multi-Scale Feature Measurement (MSFM) module.
- Employs Focal Loss during meta-training for improved performance.
Main Results:
- Achieved consistent performance gains over strong baselines on mini-ImageNet and a bearing-surface defect dataset.
- Ablation studies confirmed the contribution of each proposed module.
- Production-line tests validated the framework's practical feasibility and effectiveness.
Conclusions:
- The proposed framework significantly improves accuracy and robustness for few-shot industrial defect classification with limited data.
- Offers a practical and effective solution for real-world industrial quality control challenges.
- Demonstrates the potential of meta-learning for addressing data scarcity in specialized domains.