GMI-DETR: Insulator defect detection network based on GSConv and multi-scale isometric convolution

Read full text Summarise with VIRAT
Authors: Xuanyu Liao, Chengjiang Zhou
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

As a core component of high-voltage transmission lines, the accuracy and efficiency of insulator defect detection are directly related to the safe and stable operation of power systems. Aiming at the problems of insufficient accuracy, weak small-target defect recognition ability and poor adaptability to complex backgrounds in existing insulator detection algorithms, this study introduces a novel insulator defect detection network that integrates GSConv and multi-scale isometric convolution. Firstly, we design the Group Sparse Convolution Cross-stage Fusion Module (GSC), which reduces the number of model parameters and computational overhead while enhancing the feature discrimination ability. Secondly, the Multi-scale Isometric Convolution Module (CRMIC) is constructed to strengthen the recognition capability for small-target defects. Finally, a parallel architecture combining the CRMIC branch and the GSC branch is built to achieve efficient integration of semantic information and spatial details, thereby improving the adaptability to defects with variable scales and complex backgrounds. Experimental results show that, compared with the Baseline, the proposed GMI-DETR achieves 88.0% mAP@50-95 and 99.3% mAP@50 on public datasets, with respective improvements of 0.5% and 0.3%. The model proposed in this paper demonstrates excellent performance in surface defect detection.

Fields of Study & Subjects

Power Line Inspection Robots Infrastructure Maintenance and Monitoring Industrial Vision Systems and Defect Detection

Publication details

Journal / source
Advances in Computer Signals and Systems
Volume
10
Issue
1
Publication date
1 January 2026
Language
EN