Multiple Guidance Network for Industrial Product Surface Inspection With One Labeled Target Sample

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Traditional industrial product surface inspection methods are often data-intensive and narrowly focused, posing challenges in acquiring sufficient labeled data due to high annotation costs and limited initial production samples.
• Data scarcity hinders the development and deployment of effective automatic inspection systems, particularly for new or low-volume product categories.

Purpose of the Study:

• To propose a novel Multiple Guidance Network (MGNet) that addresses the data-hungry nature of existing industrial inspection methods.
• To enable accurate product surface inspection and object localization using minimal labeled data, specifically requiring only one support sample.

Main Methods:

• The Multiple Guidance Network (MGNet) employs a Feature Extraction Machine (FEM) to generate diverse feature maps, enhancing inspection capabilities.
• A Probability Map Generation (PMG) module facilitates coarse object positioning, while mutual and historical guidance (HG) structures leverage auxiliary datasets effectively.
• The network is designed to identify and locate specific labeled objects in query images based on a single reference sample.

Main Results:

• Comprehensive evaluations using three real-world datasets demonstrate the efficacy of the proposed MGNet.
• Experimental results confirm that MGNet can successfully perform industrial product surface inspection with only one labeled target sample.
• The method shows significant promise in overcoming data limitations in practical industrial settings.

Conclusions:

• The developed MGNet offers a data-efficient solution for industrial product surface inspection, significantly reducing the need for extensive labeled datasets.
• The network's ability to utilize a single labeled sample for identification and localization marks a substantial advancement in automated quality control.
• MGNet presents a promising and practical approach for real-world industrial applications facing data acquisition constraints.