High-Frequency Enhanced YOLOv8 for X-Ray Weld Defect Detection

Abstract

To address small scale, low contrast, weak edge texture, and edge-device deployment limitations in X-ray weld defect detection, this paper proposes High-frequency enhancement, Selective fusion, and Lightweight YOLO (HSL-YOLO). The High-Frequency Enhancement Module (HFEM) is inserted into the shallow Backbone of YOLOv8 to enhance Sobel-based horizontal and vertical edge responses, while the Hierarchical Selective Feature Pyramid Network (HS-FPN) reduces redundant Neck computation through channel attention, compression, and selective fusion. Experiments show that HFEM improves mAP50 from 0.95806 to 0.96554 and mAP50:95 from 0.54287 to 0.58475. The final HSL-YOLO achieves 0.96349 mAP50 with only 1.969M parameters, reducing parameters by about 34.6% compared with YOLOv8n. Grad-CAM++ visualization further confirms stronger attention to small defects and edge regions.