An improved RT-DETR algorithm for surface defect detection integrating state space model and feature recalibration
Industrial surface defect detection requires accurate localization of scale-varying defects under complex background interference while maintaining real-time inference efficiency. This study proposes RT-DETR-MLS, an improved RT-DETRv2-based detector integrating Mamba-based feature interaction, LSKAttention, SimAM feature recalibration, and efficient GELAN/ADown feature aggregation. Mamba is selectively incorporated into the AIFI stage and one fusion stage, while the Transformer decoder is retained for object prediction. On NEU-DET, RT-DETR-MLS achieves 80.58% mAP@50 and 161 FPS on an NVIDIA RTX 4090, compared with 73.51% and 134 FPS for RT-DETRv2. Additional comparisons with Le-DETR-M and evaluations on GC10-DET and an automotive filter dataset further support its detection performance under the investigated conditions. Controlled illumination and input-resolution experiments provide additional evidence of detection stability and accuracy–efficiency trade-offs. These findings indicate the potential of RT-DETR-MLS for real-time industrial surface defect inspection.
Authors
- 韦有江
- Wenyu Lu
- Run An
- Xiaojuan Wei
- Guofeng Wan
- Ling Diao
- Junbo Li
- Pengcheng Zhao
Institutions
- Northwest Minzu University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1038/s41598-026-81162-0
- Primary Topic
- Industrial Vision Systems and Defect Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Northwest Minzu University
- Fundamental Research Funds for the Central Universities
- Science and Technology Program of Gansu Province