The defect detection for X-ray images based on a new lightweight semantic segmentation network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Tire quality inspection relies heavily on analyzing X-ray images.
• Automated defect detection is crucial for efficient and reliable tire manufacturing.

Purpose of the Study:

• To develop an end-to-end lightweight semantic segmentation network for automated bead toe error detection in tire X-ray images.
• To introduce a novel evaluation metric, local mean Intersection over Union (L-mIoU), for assessing segmentation performance.

Main Methods:

• An encoder-decoder architecture was employed to extract and fuse texture features from tire X-ray images.
• The network was trained to segment the bead toe region accurately.
• A new metric, L-mIoU, was proposed alongside the standard mIoU for evaluation.

Main Results:

• The proposed network achieved a segmentation accuracy of 97.1% mIoU on a tire X-ray dataset.
• The novel L-mIoU metric recorded a score of 92.4% for the segmentation effect.
• The system calculated bead toe coordinates in just 1.0 second for 512x512 images.

Conclusions:

• The lightweight semantic segmentation network effectively detects bead toe errors in tire X-ray images.
• The proposed L-mIoU metric provides a valuable assessment of segmentation quality.
• The method offers a fast and accurate solution for automated tire quality inspection.