Abstract
<title>Abstract</title> <p>Metal surface defect detection faces significant hurdles in modern intelligent manufacturing. Traditional image processing algorithms often struggle to identify complex and irregular defects, while relying solely on artificial intelligence algorithms face difficulties in achieving effective image preprocessing. This paper aims to explore an algorithm that integrates wavelet transform into deep vision to enhance detail features. Through the frequency domain decomposition of wavelet transform and attention mechanisms, high-frequency details are explicitly preserved and enhanced, making this approach particularly suitable for tasks requiring fine boundaries, such as object detection and image segmentation. The Haar wavelet achieves low computational complexity by using only integer addition and subtraction and can effectively improve model accuracy without sacrificing efficiency. The “dual-stream wavelet” is an important extension of the aforementioned Haar wavelet module, primarily reflected in two aspects: network architecture design and frequency-spatial feature fusion. While traditional wavelet modules only use Discrete Wavelet Transform Module (DWT) as a fixed preprocessing or down-sampling tool, the dual-stream wavelet significantly enhances feature representation capability through parallel architecture and learnable fusion. The Wavelet-Conv path effectively preserves details such as edges and textures by explicitly decomposing the image into LL, LH, HL, and HH sub-bands and using learnable weights to boost high-frequency components. Moreover, the standard down-sampling path learns abstract semantic features and fuses them with the former via concatenation, enabling complementary information exchange. In the dual-stream framework, the transformation from fixed wavelet bases and simple learnable weights to end-to-end adaptive fusion of two streams results in a frequency-aware, multi-receptive-field down-sampling module. Particularly, the High-Freq-Enhance module enhances high-frequency details in feature maps through residual learning. The residual connection ensures the fidelity of the original information, while the attention mechanism determines, based on global information, which channels require high-frequency enhancement. Placed in the deeper layers of the backbone, it compensates for high-frequency details attenuated by repeated down-sampling. The core insight of this architecture is that down-sampling in traditional CNNs is "discard-based", whereas wavelet-enabled down-sampling is "encoding-based" and "recoverable". Compared to standard YOLO, this approach proposes three key improvements: (1) High‑frequency information retention: Conventional down-sampling discards high‑frequency details, whereas the dual‑stream wavelet transform encodes and preserves them.(2) Multi‑path feature fusion: Standard CNNs rely solely on single‑path spatial features, while the dual‑stream design achieves effective multi‑path information integration.(3) Small‑target enhancement: After repeated down-sampling, small‑target signals are prone to being overwhelmed by noise; the proposed high‑frequency enhancement module selectively restores detailed responses to mitigate this issue.</p>