物探化探计算技术2025，Vol.47Issue(4)：566-576,11.DOI:10.12474/wthtjs.20240423-0001
用于探地雷达实测道路异常解译数据的小样本增强方法
Small-sample augmentation method for interpretation data of road anomalies measured by ground penetrating radar
摘要
Abstract
In order to obtain the underground anomaly information of roads with diverse categories and complex features,ground-penetrating radar(GPR)used for road anomaly detection produces a large amount of image data.However,the current interpretation of the data mainly dependents on manual methods,resulting in very low interpretation efficiency.Extending the deep learning(DL)method,which is widely used for target recognition,to GPR real data decoding can effectively overcome the shortcomings of the traditional manual approach.However,the DL method requires a large amount of successfully deciphered data as a training set,and the need for confidentiality,information silos,and other real-world constraints seriously limit the collection of GPR deciphered data.Aiming to address the issue that the scarcity of GPR decoding data hinders DL model training,this paper proposes a small-sample-oriented self-supervised conditional generative adversarial network(SS-CGAN)method for generating GPR road anomaly decoding data,thereby expanding the decoding dataset and forming a training set for the DL model.The experimental results show that the data generated by the SS-CGAN small-sample enhancement method proposed in this paper is in good agreement with the features of the GPR measured data,and the supplementation of the generated data to the training set of the DL-based Faster RCNN model increases the average recall rate of the trained model on the GPR measured data by 6.87%,and an increase of the average accuracy rate by 11.56%.关键词
探地雷达/道路/数据增强/生成对抗网络/小样本Key words
ground penetrating radar/road/data augmentation/GAN/small-sample分类
信息技术与安全科学引用本文复制引用
王娇,陈宣,姜彦南..用于探地雷达实测道路异常解译数据的小样本增强方法[J].物探化探计算技术,2025,47(4):566-576,11.基金项目
国家自然科学基金(62371147) （62371147）
电波环境特性及模化技术重点实验室基金(202003007) （202003007）
广西无线宽带通信与信号处理重点实验室基金(GXKL06200126) （GXKL06200126）
