YOLOv11-AHFE: adaptive high-frequency signal enhancement mechanism for surface detection

Metal surface defect detection confronts persistent challenges in modern intelligent manufacturing, where traditional image processing struggles with complex and irregular defects, while the feature extraction capabilities of standard of deep learning models is insufficient. To solve these limitations, this paper proposes an advanced detection framework that integrates wavelet transforms into deep vision architectures to enhance fine-grained detail features representation. Different from conventional convolutional networks where down-sampling is strictly “discard-based”, our strategy introduces a dual-stream wavelet module characterized by “encoding-based” and recoverable transformations. This design features a parallel Wavelet-Conv path that decomposes images into multi-frequency sub-bands (LL, LH, HL, and HH) to explicitly preserve edges and textures, alongside a standard path that captures abstract semantic features. Through end-to-end adaptive feature fusion, the two streams achieve complementary information exchange. Furthermore, a dedicated High-Frequency Enhancement module is integrated into the deeper layers of the backbone. By combining residual learning with attention mechanisms, it selectively compensates for high-frequency details that are typically attenuated by repeated down-sampling while suppressing illumination-induced noise. Compared to conventional baselines, this architecture delivers three distinct advantages: (1) High-frequency feature preservation: Explicitly encodes and retains critical spatial details during down-sampling, mitigating the inherent information loss of standard convolutions. (2) Dual-domain feature fusion: Integrates parallel frequency and spatial pathways to construct a highly robust multi-scale representation. (3) Micro-defect enhancement and noise suppression: selectively restores fine-grained feature responses to prevent subtle defects from being overwhelmed, demonstrating high-reliably robustness against environmental disturbances such as illumination fluctuations and sensor noise.