YOLO algorithm with hybrid attention feature pyramid network for solder joint defect detection

Li Ang,Siti Khatijah Nor Abdul Rahim,Raseeda Hamzah,Raihah Aminuddin,Gao Yousheng

2024-01-02

Abstract:Traditional manual detection for solder joint defect is no longer applied during industrial production due to low efficiency, inconsistent evaluation, high cost and lack of real-time data. A new approach has been proposed to address the issues of low accuracy, high false detection rates and computational cost of solder joint defect detection in surface mount technology of industrial scenarios. The proposed solution is a hybrid attention mechanism designed specifically for the solder joint defect detection algorithm to improve quality control in the manufacturing process by increasing the accuracy while reducing the computational cost. The hybrid attention mechanism comprises a proposed enhanced multi-head self-attention and coordinate attention mechanisms increase the ability of attention networks to perceive contextual information and enhances the utilization range of network features. The coordinate attention mechanism enhances the connection between different channels and reduces location information loss. The hybrid attention mechanism enhances the capability of the network to perceive long-distance position information and learn local features. The improved algorithm model has good detection ability for solder joint defect detection, with mAP reaching 91.5%, 4.3% higher than the You Only Look Once version 5 algorithm and better than other comparative algorithms. Compared to other versions, mean Average Precision, Precision, Recall, and Frame per Seconds indicators have also improved. The improvement of detection accuracy can be achieved while meeting real-time detection requirements.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the issues in solder joint defect detection in Surface Mount Technology (SMT), where traditional manual inspection methods are inefficient, inconsistent in evaluation, costly, and lack real-time data. Although existing automated detection methods have improved detection to some extent, they still suffer from low detection accuracy, high false alarm rates, and high computational costs. Particularly, the detection of small-sized defects remains ineffective with current methods.

To tackle these problems, the authors propose a new method, namely the Hybrid Attention Feature Pyramid Network (HA-FPN), and apply it to the YOLOv5 detection model. Specifically, this method aims to:

1. **Improve detection accuracy**: Enhance the network's ability to perceive contextual information by using Enhanced Multi-Head Self-Attention (EMSA) and Coordinate Attention (CA) mechanisms, thereby improving detection accuracy.
2. **Reduce computational cost**: Optimize the feature fusion method to reduce the loss of shallow feature information, improve the network's ability to perceive long-distance positional information, and enhance the learning ability of local features.
3. **Achieve real-time detection**: Ensure real-time detection requirements in industrial production environments while maintaining detection accuracy.

Through these improvements, the paper hopes to significantly enhance the accuracy and efficiency of solder joint defect detection, especially in detecting small-sized defects.