Abstract

Weld defect detection is challenged by significant variations in defect scale, complex defect morphology, and strong background interference. To address these issues, this study proposes WID-YOLO11, an improved detection model based on YOLO11n. First, a novel C3k2_LFA-WTConv module is developed to decompose feature maps into frequency subbands via learnable wavelet filters and adaptively reweight them through a Learnable Frequency-Attention (LFA) mechanism, enabling the network to learn task-adapted subband representations for different defect types, where each class exhibits distinct wavelet subband energy distributions. Second, an Improved Multi-Scale Dilated Local Attention (IMSDA) module is constructed by extending the dilation-rate set of the original MSDA from three to four values (1, 2, 3, and 4), expanding the maximum equivalent receptive field from 7 × 7 to 9 × 9 to improve multi-scale sensitivity to medium-scale defects such as porosity and spatter at low computational overhead. Finally, a DySample dynamic upsampling module is integrated into the neck network to adaptively learn sampling locations and preserve defect boundary details more effectively during multi-scale feature fusion. Evaluations on a publicly available weld defect dataset of 986 images show that WID-YOLO11 outperforms the YOLO11n baseline by 4.5, 10.4, and 3.4 percentage points in mAP@0.5, Recall, and mAP@0.5:0.95, respectively, confirming the effectiveness of the three proposed enhancements.

IPC Classification

Keywords