SteelSpec-RTDETR: Defect-Aware Multi-Scale Transformer Detection for Low-Contrast Steel Surface Inspection

Abstract

: Accurate and efficient detection of steel surface defects is essential for automated industrial quality inspection. However, steel defects often exhibit small sizes, low contrast, irregular shapes, and strong similarity to background textures, making them difficult to localize reliably under real-time deployment constraints. To address these challenges, this paper proposes SteelSpec-RTDETR, an accuracy-efficiency-balanced multi-scale Transformer detector for industrial steel surface defect detection. The proposed model is developed based on the end-to-end detection paradigm of RT-DETR, while redesigning the feature extraction and feature fusion stages to better capture defect-sensitive representations. Specifically, a defect-oriented backbone is introduced, consisting of a Surface Background Smoothing Block, Local-Spectral Defect Mixing Blocks, and an Agent-based Defect Context Attention module. These components suppress background-dominant surface responses, enhance local and high-frequency defect cues, and capture compact long-range contextual dependencies. Furthermore, a Defect-Aware BiFPN and an Intra-Level Defect Aggregation module are designed to improve cross-scale feature interaction and adaptively integrate complementary representations from different fusion depths. The refined multi-scale features are then used as memory tokens for uncertainty-minimal query selection and decoding, enabling accurate end-to-end detection without non-maximum suppression. Experiments on the NEU-DET and GC10-DET datasets demonstrate that SteelSpec-RTDETR achieves a favorable balance between detection accuracy and computational cost. On NEU-DET, the proposed model achieves 81.6% mAP, 88.5% mAP 50 , and 74.2% mAP 75 . On GC10-DET, it obtains 76.4% mAP, 88.4% mAP 50 , and 68.1% mAP 75 . The complete model contains 35.6 M parameters, requires 59.4 G FLOPs, and achieves an inference latency of 22.7 ms.