Reference-Based Defect Detection Network.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Defect detection is crucial in industrial settings, often framed as an object detection task.
• Existing object detectors face challenges like texture shift and partial visual confusion, limiting their effectiveness.
• Texture shift occurs when models encounter unseen textures, while partial visual confusion arises from similarities between partial and complete defect features.

Purpose of the Study:

• To develop an advanced defect detection network that overcomes limitations of standard object detectors.
• To introduce novel reference mechanisms to enhance robustness against texture variations and visual ambiguities.
• To improve the accuracy of defect classification and bounding box regression in industrial applications.

Main Methods:

• Proposed the Reference-based Defect Detection Network (RDDN) incorporating template and context references.
• Template reference was used to mitigate texture shift at image, feature, or region levels, focusing detectors on defective areas.
• Context reference leveraged surrounding information from larger bounding boxes to refine region classification and regression.

Main Results:

• The RDDN effectively addressed the challenges of texture shift and partial visual confusion in defect detection.
• Template references, using either actual or pseudo templates, were jointly trained with detectors using normal sample supervision.
• Context references improved the precision of classifying and localizing defects by utilizing contextual information.

Conclusions:

• The proposed RDDN significantly enhances defect detection performance in industrial computer vision.
• The integration of template and context references offers a robust solution to common challenges in defect identification.
• Experimental validation on two datasets confirms the superiority of the RDDN approach.