OHEH-RTDETR: An Improved RT-DETR Detection Model Based on Frequency Layered Processing and Advanced Feature Selection

Abstract As a fundamental material in modern industry, steel finds extensive application across various sectors, including manufacturing, construction, and energy. Steel product surface defects exhibit characteristics like multiple types, scales, small targets, and minimal background differences. Small object defects are challenging to detect due to their small image resolution and sparse feature information. To enable accurate and fast detection of industrial defects, this paper proposes an improved Real-Time-Detection-Transformer (RT-DETR)-based defect detection method that integrates high and low-frequency information processing and efficient advanced-feature-based selection and fusion, aiming to enhance the effectiveness of detecting multi-scale small targets. By leveraging contextual information and attention mechanisms, the method employs orthogonal attention-based deep feature extraction and a high-low frequency layered processing framework to select and fuse advanced features. It enriches extracting and integrating relationships between high- and low-level defect features by identifying spatial pixel-level relationships. The proposed algorithm achieves a mean average precision (mAP) of 91.8% and a detection speed of 135.6 FPS, meeting the demands of real-time industrial detection and achieving a balance between detection accuracy and detection speed. Generalization experiments on the public NEU-DET and GC10-DET datasets indicate mAP50 improvements of 2.9% and 9.1%, respectively, and the enhanced algorithm boosts recall rates for most small defect types, especially with a 15.3% increase in the recall rate for irregular crack defect. These results demonstrate that OEHE-RTDETR holds promise for industrial real-time detection.