High-Frequency Guided Feature Aggregation for Strip Steel Surface Defect Detection

Abstract

Strip steel surface defect detection (S3D) is a critical task in industrial quality control, and existing detectors often struggle to precisely localize fine defect structures within complex textured backgrounds. High-frequency cues, such as subtle edges and texture patterns, are known to be effective for capturing fine details; however, they are often progressively attenuated or even lost in deep convolutional backbones due to their intrinsic low-pass filtering behavior. To address this limitation, we propose a high-frequency guided feature aggregation (HFFA) framework that explicitly decouples low-frequency semantic information and high-frequency defect-sensitive cues into two complementary representation pathways. Specifically, the backbone network is explicitly interpreted as a low-frequency semantic extractor for modeling global context and defect semantics, while a Laplacian feature pyramid network (LFPN) is introduced in parallel to explicitly extract multiscale high-frequency structural features by removing low-frequency components. Based on this frequency-decoupled representation, we design high-frequency injection modules (HFIMs) that use fixed high-frequency priors to modulate backbone features, selectively enhancing defect-relevant regions while suppressing noise. Finally, we develop a routing enhancement module (REM) that enable content-adaptive multiscale collaboration through dynamic spatial routing, improving the adaptability of fused representations to varying defect scales. Experimental evaluation on three benchmark datasets for S3D shows that HFFA consistently outperforms twenty-three SOTA methods. Furthermore, cross-domain evaluation on pipeline weld defect detection confirms its generalization, while integration of its key high-frequency extraction and injection components into the YOLO series substantially improves their prediction accuracy, confirming the robustness and transferability of our approach. Our code is available at https://github.com/hpguo1982/HFFA Note to Practitioners—The motivation for this paper is to develop a high-accuracy visual inspection method for the surface defect detection of strip steel. Existing approaches often struggle to capture subtle edges and texture patterns within complex surfaces, limiting their ability to localize fine defects. To address the issue, we propose an HFFA framework that integrates Laplacian enhancement, high-frequency injection, and adaptive routing to explicitly extract and fuse high-frequency cues, enabling more precise and stable representations. As a result, HFFA achieves superior detection accuracy on three public surface-defect benchmarks, as well as in cross-domain and cross-model evaluations, confirming its robustness, generalization, and transferability. Therefore, HFFA can be readily applied to a variety of industrial surface inspection tasks and supports the development of intelligent quality-assurance systems in manufacturing.