Abstract
Robust defect detection is a crucial link in vision-based measurement systems. However, noise and lighting in complex industrial environments can weaken the structural and edge information of defects, leading to unsatisfactory measurement accuracy. Recent studies have employed discrete wavelet transform (DWT) to enhance defect features using multi-resolution frequency subbands. However, they exhibit a bias toward frequency branches that compromise spatial cues, resulting in inadequate feature enhancement. To address these challenges, this study considers enhancing measurement features from a joint spatial–frequency perspective and proposes a wavelet-guided spatial-frequency prior network (WSP-Net). At the structural level, a dynamic prior excavator and a dual-domain detail aggregator are constructed to adaptively mine and aggregate discriminative spatial-frequency defect priors. At the decision level, a frequency-assisted task-aware head (FT-Head) is customized to further enhance task-specific semantics and edges through decoupling DWT. In addition, joint supervision is introduced to constrain fine regression for complex defects with ambiguous boundaries. Extensive experiments demonstrate that the proposed method achieves state-of-the-art (SOTA) performance across various vision-based measurement systems (e.g., PCB: 95.7% mAP; steel plates: 82.1% mAP; pipelines: 98.7% mAP), while maintaining low computational cost (Parameters: 9.1M, FLOPs: 12.8G FPS: 235-284 imgs/s). It also achieves remarkable performance under strong noise and low-light conditions. Furthermore, generalization analysis using real-world pipeline data from MFL-based measurements indicates that the method holds strong potential for engineering applications in real-world measurement systems.
Funding Information
  • National Natural Science Foundation of China (U21A20481)
  • National Natural Science Foundation of China (62373085)
  • Fundamental Research Funds for the Central Universities of China (N2404014)

This publication has 29 references indexed in Scilit: