Karst formations, prevalent in western mountainous regions, pose significant challenges to construction projects. Ground-penetrating radar (GPR) offers a non-destructive means of detecting these unfavorable geological conditions within tunnels. However, GPR image interpretation is often subjective and limited by human expertise. To address this, this study proposes a novel approach that combines GPR interpretation feature patterns and a deep learning-based target detection model (YOLOv4). By analyzing GPR data based on amplitude, frequency, phase, and two-way travel time, eight distinctive patterns were identified for karst-related features. These patterns were used to calibrate existing data, enabling more accurate and consistent interpretations. The efficacy of the proposed method was evaluated using GPR data from the Jiangyou section of the JiuMian Expressway. Results demonstrate that the deep learning model effectively interprets GPR data, achieving a mean average precision (mAP) of 45.12%, thereby meeting the requirements of engineering construction The proposed method provides a new idea for using deep learning for intelligent interpretation of GPR.