DCF-DETR: Collaborative Context and Cross-Level Feature Calibration for Steel Defect Detection

Read full text Summarise with VIRAT
Authors: Yilu Hao, Yuzhou Liu, Shu Gong, Yin Chen, Suping Liu
Cited by: 0
References: 38

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

Detecting surface defects on structural steel remains challenging due to low contrast, multi-scale variations, and severe class imbalance. To tackle these issues, we present DCF-DETR, a tailored end-to-end detection framework for robust industrial defect inspection. Our framework effectively integrates several key designs: a lightweight Collaborative Context System (CCS) that employs parallel standard and dilated convolutions to capture multi-scale contextual cues, enhancing the recognition of texture-similar defects; and a Cross-level Feature Calibration Module (CFCM), which enables bidirectional feature interaction across network layers using a local-global attention strategy to bridge semantic gaps. Furthermore, we adopt and adapt the FMPDIOU loss, which incorporates focal modulation to accelerate convergence under sample imbalance, along with Dysample, a dynamic upsampling mechanism that adaptively restores spatial details for finer defect localization. Extensive experiments on the NEU-DET and GC10-DET datasets show that DCF-DETR achieves strong performance—reaching 79.1% and 71.5% mAP50, respectively—while maintaining high inference speed (73 FPS), highlighting its effectiveness as a practical solution for real-time industrial quality inspection.

Fields of Study & Subjects

Industrial Vision Systems and Defect Detection Infrastructure Maintenance and Monitoring Non-Destructive Testing Techniques

Publication details

Journal / source
IEEE Access
Volume
14
Pages
50057–50073
Publication date
1 January 2026
Language
EN