Research on Intelligent Recognition Algorithms for Common Subgrade Diseases Based on Radar Images

China's operational roadbed diseases show a growing trend in terms of both diversity and quantity, leading to an increasing frequency of road safety accidents caused by these diseases. It is of significant importance to identify the location, morphology, and developmental degree of hidden roadbed diseases and address them accordingly for the safe operation and maintenance of highways during their operational period. Current non-destructive testing methods such as ground-penetrating radar heavily rely on subjective experience for data processing and interpretation, which no longer meet the growing demand for precise and rapid identification of disasters in roadbed engineering. In this study, we established a training set and a test set using images of common roadbed diseases, specifically roadbed looseness and roadbed voids, in a 1:4 ratio, and labeled the disease types in each image. We improved the Faster R-CNN algorithm and obtained two enhanced versions, namely faster_rcnn_inception_v2 and faster_rcnn_resnet50. Both algorithms were trained and indicators such as loss value, test recognition area, and accuracy were analyzed. The results showed that the faster_rcnn_inception_v2 algorithm outperformed the faster_rcnn_resnet50 algorithm with a total loss value of 0.0235, a total of 425 recognized regions, and an accuracy of 91.1% for region recognition. Therefore, the faster_rcnn_inception_v2 algorithm is chosen for the image detection and recognition of roadbed diseases in urban roads.