Abstract: Three-dimensional ground penetrating radar (GPR) technology is a critical method for detecting latent road defects. Analyzing GPR data with intelligent defect recognition algorithms can enhance detection efficiency. However, existing algorithms have not fully utilized the multi-view information provided by GPR, resulting in lower recognition accuracy. To address this issue, this paper proposes a GPR multi-view fusion model for subgrade defect identification. The model employs a dual-channel parallel structure, utilizing the MobileViT network to extract high-dimensional features from both B-scan and C-scan images of GPR data, thereby tapping into the complementary information inherent in the dual views for feature learning. To effectively integrate information from these dual views, a feature interlaced weighted fusion method is proposed, arranging the high-dimensional features extracted from the dual views in proximity and dynamically optimizing the weight allocation of feature channels to highlight key features while suppressing irrelevant information. Experimental results demonstrate that the network model achieves an accuracy rate of 90.5% on the test set. Moreover, under Gaussian white noise interference, compared with baseline models, the overall degradation index of the model decreases by 13.51%, showcasing superior robustness.

Key words: disease detection, deep learning, ground penetrating radar, multi-view fusion