A welding defect detection model based on a shape-aware network

Because X-ray welding images have complex backgrounds and welding defects have different sizes and shapes, effectively detecting welding defects in X-ray images is still a challenge. To solve these problems, a shape-aware network (SA-NET) was proposed, whose core was the shape-aware module (SAM). SA...

Full description

Saved in:
Bibliographic Details
Published in:Journal of intelligent & fuzzy systems 2022-01, Vol.43 (5), p.6147
Main Authors: Shi, Lukui, Zu, Haoran, Tai, Jikai, Niu, Weifei
Format: Article
Language:English
Subjects:
Boxes
Classification
Feature maps
Regression
Shape recognition
Weld defects
Welding
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Because X-ray welding images have complex backgrounds and welding defects have different sizes and shapes, effectively detecting welding defects in X-ray images is still a challenge. To solve these problems, a shape-aware network (SA-NET) was proposed, whose core was the shape-aware module (SAM). SAM includes a free-shape region proposal network (FS-RPN) and a two-level regression head (TR-Head). FS-RPN predicts the shape of the anchor boxes corresponding to each position on the feature maps, and aligns the feature maps according to the predicted anchor box shape. Then, the offset and the foreground classification score of the anchor boxes are predicted according to the aligned feature maps. Thus, FS-RPN generates the proposal regions with a higher quality. TR-Head uses the first-level detection head, which only contains one regression branch, to further improve the quality of the proposal regions by fine-tuning the proposal regions. It employs the second-level detection head, which consists of one classification branch and one regression branch, to predict the categories and the boxes of defects. The experimental results showed that SA-NET effectively improved the quality of the proposal regions and greatly improved the detection effect of welding defects, especially defects with special shapes.
ISSN:1064-1246
1875-8967
DOI:10.3233/JIFS-220132