Abstract: Aiming at the problems of low accuracy, false detection and missed detection in ground penetrating radar (GPR) image detection of underground pipelines in substations, an improved detection model CDG-YOLOv13n is proposed in this paper. The model embeds the CBAM attention mechanism in the Backbone part to strengthen the perception of target features and key areas. The Neck part uses the dynamic upsampling module DySample to replace the traditional method to improve the performance of detail preservation and multi-scale feature fusion. At the same time, the bounding box regression loss function is adjusted from CIoU to GIoU, which effectively improves the positioning accuracy when the prediction box and the real box are not aligned. Experiments show that the accuracy (P) of the proposed model on the measured GPR data set is 98.1%, the recall rate (R) is 95.5%, and the mAP50 and mAP50-95(%) are 98.4% and 78.7%, respectively, which are 1.3%, 0.2%, 1.5% and 1.2% higher than the original YOLOv13n model. Compared with Faster R-CNN, SSD, YOLOv5n and YOLOv8n, mAP50 increased by 6.2%, 12.2%, 6.5% and 3.6%, respectively. In addition, a professional interface program is developed to realize automatic call and integration with the ground penetrating radar data acquisition software Multi-Gprview, which can support real-time processing and recognition on site and meet the engineering requirements of rapid and mobile detection of underground pipelines in old substations.