Related Experiment Video
Updated: Jan 15, 2026
05:47
Simulation of a Scaled Assembly Process with Collaboration of a Robotic Arm and Monitoring through a Vision System for Quality Control
Published on: August 29, 2025
417
Contrastive learning enhanced MobileMamba for real time industrial defect detection on edge devices
Jun Huang1,2, Shamsul Arrieya Ariffin3,4, Qun Yang5
1Faculty of Intelligent Manufacturing, Wuhu Vocational Technical University, Wuhu, 241006, China. huangjun1207@whit.edu.cn.
Scientific Reports
|January 13, 2026
Summary
This study introduces a novel lightweight Mamba-based network for real-time industrial metal defect detection on edge devices. The efficient nano model achieves high accuracy, outperforming existing methods.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Materials Science
Background:
- Real-time industrial metal defect detection is computationally intensive for edge devices.
- Existing methods struggle with limited computing power and real-time processing requirements.
Purpose of the Study:
- To develop a lightweight, efficient, and accurate defect detection network for industrial applications on edge devices.
- To leverage Mamba's capabilities for enhanced feature extraction and real-time performance.
Main Methods:
- Designed a MobileMamba backbone network incorporating Mamba for temporal and global feature focus.
- Employed a multi-scale approach to enhance Mamba's feature extraction.
- Implemented a contrastive auxiliary loss function with difficult/easy negative sample mining.
Main Results:
- The proposed model demonstrates superior efficiency compared to same-level detection models.
- Achieved real-time reasoning on edge devices with higher accuracy than comparable models.
- Validated on NEU-DET, GC10-DET, and APDDD datasets.
Conclusions:
- The lightweight Mamba-based network effectively addresses real-time industrial defect detection challenges on edge devices.
- The model offers a balance of high efficiency and accuracy for practical industrial deployment.
- This approach enables advanced defect detection capabilities on resource-constrained hardware.
