TGDNet: A Multi-Scale Feature Fusion Defect Detection Method for Transparent Industrial Headlight Glass

Area of Science:

• Computer Vision
• Industrial Automation
• Materials Science

Background:

• Automotive headlight lens defect detection is crucial but challenging due to varied defect types and transparency.
• Existing methods struggle with identifying subtle, transparent surface defects.

Purpose of the Study:

• To develop an accurate and efficient computer vision solution for transparent glass defect detection in automotive headlights.
• To introduce a novel deep learning network and dataset for this specific application.

Main Methods:

• Collected 2000 automotive headlight images, categorized defects (spots, scratches, abrasions), and applied a dataset augmentation method (SWAM) to create the Lens Defect Dataset (LDD) of 5532 images.
• Proposed the Transparent Glass Defect Network (TGDNet) featuring a Transparent Glass Feature Extraction (TGFE) module and improved attention mechanisms for multi-scale feature fusion.
• Utilized multi-angle lighting during data acquisition.

Main Results:

• TGDNet demonstrated superior performance on the LDD compared to classical defect detection methods.
• Achieved a 6.7% improvement in mean Average Precision (mAP) and an 8.9% improvement in mAP50 over the best baseline.
• The SWAM augmentation method effectively addressed class imbalance in the dataset.

Conclusions:

• The proposed TGDNet, leveraging multi-angle lighting and advanced feature extraction/fusion, offers a robust solution for transparent glass defect detection.
• The LDD and TGDNet provide valuable resources for advancing research in industrial inspection of transparent materials.