Textured Surface Defect Detection with Image-Level Accuracy Optimization via Patch-Level Features and Simulation-Based Samples

Textured surface defect detection plays a critical role in industrial product quality monitoring. However, excessive reliance on pixel-level precision, especially when defect boundaries are unclear, can lead to more complex models, lower efficiency, and increased risk of false positives or missed de...

Full description

Saved in:
Bibliographic Details
Published in:2025 4th Asia Conference on Algorithms, Computing and Machine Learning (CACML) pp. 1 - 7
Main Authors: Li, Fen, Li, Dongxiao, Huang, Le, Chen, Hangrui, Zhang, Ming
Format: Conference Proceeding
Language:English
Published: IEEE 28.03.2025
Subjects:
Accuracy
Computational efficiency
Computational modeling
Costs
Defect detection
Feature extraction
Labeling
Monitoring
Patch-Level Features
Robustness
Surface Defect Detection
Surface texture
Synthetic Defects
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Textured surface defect detection plays a critical role in industrial product quality monitoring. However, excessive reliance on pixel-level precision, especially when defect boundaries are unclear, can lead to more complex models, lower efficiency, and increased risk of false positives or missed detections, ultimately compromising system reliability. Additionally, challenges such as the difficulty of collecting defect samples, severe class imbalances, and high pixel-level labeling costs hinder the effectiveness of convolutional neural networks in terms of both inference speed and computational efficiency. To overcome these challenges, we propose SiamSimNet, a dual-channel Siamese network that relies on high-quality training data generated through simulation algorithms, minimizing the need for real defect samples. By incorporating locally aware patch features, the proposed method significantly improves detection accuracy and robustness, while maintaining high inference efficiency. This approach shows strong potential for real-world industrial applications.
DOI:10.1109/CACML64929.2025.11010951