An automatic recognition algorithm for GPR images of RC structure voids

• A novel method based on SVM is proposed to automatically identify voids in GPR images.
• The trained SVM model achieves high accuracy in void depth and lateral range location.
• Statistic-based feature extraction approach of GPR images is concise and informative.
• Predictive deconvolution is used to suppress multiples and improve SNR of GPR images.

Ground penetrating radar (GPR) is a powerful tool for detecting defects in and behind reinforced concrete (RC) structures. However, the traditional way of interpreting GPR data involves considerable manpower and is time-consuming. The aim of this study is to illustrate a new approach to recognize GPR images of RC structure voids automatically. Firstly, synthetic GPR images are created by FDTD method. As multiple waves caused by steel bars seriously interfere with the target echo signals, it is difficult to identify targets from the forward modeling images. According to the periodicity of multiple waves from steel bars, the predictive deconvolution method is used to suppress those waves and the outcome is preferable. Then, the support vector machine (SVM) algorithm is proposed to automatically recognize voids in GPR images. The automatic identification procedure includes four steps: 1) collecting training data, 2) extracting features from GPR images, 3) building the SVM model and 4) identifying the voids automatically. The results show that the proposed method provides a suitable tool to locate the cover depths and lateral ranges of the voids, and the trained SVM model gives a favorable outcome when noise (no more than 5%) is added to a synthetic GPR image.