Surface defect detection method for electronic panels based on attention mechanism and dual detection heads
Le Wang1, Xixia Huang1, Zhangjing Zheng1
1Institute of Logistics Science and Engineering, Shanghai Maritime University, Shanghai, People's Republic of China.
Plos One
|January 13, 2023
Summary
This study introduces a novel deep learning approach for automatic electronic panel surface defect detection. The method enhances feature extraction for irregular and small defects, significantly improving detection accuracy over existing techniques.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Manufacturing Quality Control
Background:
- Traditional methods for electronic panel surface defect detection are limited by their inability to handle irregular shapes and small targets, restricting their application in real production.
- Existing techniques lack robustness and struggle with feature extraction for complex defect characteristics, necessitating advanced solutions.
Purpose of the Study:
- To develop a novel deep learning-based target detection technique for electronic panel surface defect detection.
- To enhance the interpretability and performance of deep learning models for detecting irregular and small surface defects in electronic panels.
Main Methods:
- Application of deep learning target detection techniques to electronic panel surface defect analysis.
- Design of a deformable convolution module with a convolutional self-attention module to learn offsets for irregular defect features.
- Implementation of a dual detection head incorporating the Squeeze-and-Excitation (SE) mechanism to address small target characteristics.
Main Results:
- The proposed method achieved a map_0.5 metric of 78.257% on a custom electronic panel defect dataset.
- Demonstrated a significant improvement over existing methods, with a 13.506% increase compared to YOLOV5 and a 33.457% increase compared to Retinanet.
- Experimental results confirmed enhanced interpretability and superior metric performance of the proposed approach.
Conclusions:
- The novel deep learning method effectively addresses the challenges of detecting irregular and small surface defects in electronic panels.
- The integration of deformable convolutions, self-attention, and SE mechanisms provides a robust and interpretable solution for automated quality control.
- The proposed approach shows significant potential for large-scale application in real-world electronic panel production environments.
