The manual identification of non-metallic pipelines in ground-penetrating radar (GPR) images is inefficient and heavily experience-dependent. Existing deep-learning methods suffer from performance degradation caused by image noise, signal attenuation, and false anomalies. To address these challenges, this paper proposes YOLOv8-WT, which introduces a novel WTConv-ATT module combining wavelet transform and multi-dimensional dynamic attention. This module performs multi-level wavelet decomposition to extract frequency-domain features and enhance global and low-frequency information perception; meanwhile spatial-channel-pixel attention adaptively generates feature fusion weights to suppress noise interference. Several existing well-established modules (Wise-IoU, C2f-FSDA, CBAM, EMA) are also integrated to further boost detection performance. Experimental results on merged public GPR datasets show that compared with the YOLOv8s baseline, the proposed model achieves increases of 1.27, 4.59, and 2.55 percentage points in Precision, Recall, and mAP50, reaching 92.98%, 91.83%, and 96.76%, respectively. YOLOv8-WT obtains promising detection performance for non-metallic pipelines under mixed-dataset conditions. Further validation is still required for unknown real-world GPR scenarios.