A Lightweight Deep Network for Defect Detection of Insert Molding Based on X-ray Imaging

Area of Science:

• Industrial automation
• Non-destructive testing
• Deep learning for quality control

Background:

• Insert molding processes require robust defect detection for quality assurance.
• Traditional methods struggle with detecting tiny defects and variations in X-ray image quality.
• Existing deep learning models can be computationally intensive for industrial implementation.

Purpose of the Study:

• To develop a lightweight and effective deep network for abnormality detection in industrial insert molding processes using X-ray images.
• To enhance the detection of tiny defects and improve robustness against image variations.
• To create a practical solution for real-time industrial inspection.

Main Methods:

• A novel deep network based on the YOLOv5 architecture was developed.
• Fast guide filtering was applied to digital radiography (DR) images.
• A multi-task detection dataset using overlap slices was constructed for tiny target detection.
• Embedded Ghost modules and transformer modules were incorporated for model lightening and feature extraction.

Main Results:

• The proposed method achieved a mean Average Precision (mAP) of 93.6%, outperforming peer networks by 3%.
• The network demonstrated robustness against luminance variations and blurred noise.
• Ablation studies confirmed a 32% model size reduction due to the Ghost module.
• The method proved effective in detecting tiny targets and improving overall DR defect detection.

Conclusions:

• The developed lightweight deep network offers a highly accurate and robust solution for insert molding process inspection.
• The integration of Ghost and transformer modules significantly enhances detection performance and model efficiency.
• The proposed method provides a valuable reference for automated quality control in industrial manufacturing.