Ground penetrating radar (GPR) has been widely used as a non-destructive technique to detect subsurface objects. Manual interpretation of GPR data is tedious and time-consuming. To address this challenge, an automatic method based on a deep instance segmentation framework is developed to detect and segment object signatures from GPR scans. The proposed method develops the Mask Scoring R-CNN (MS R-CNN) architecture by introducing a novel anchoring scheme. By analyzing the characteristics of the hyperbolic signatures of subsurface objects in GPR scans, a set of anchor shape ratios are optimized and selected to substitute the predefined and fixed aspect ratios in the MS R-CNN framework to improve the signature detection performance. In addition, transfer learning technique is adopted to obtain a pre-trained model to address the challenge of insufficient GPR dataset for model training. The detected and segmented signatures can then be further processed for target localization and characterization. GPR data of tree roots were collected in the field to validate the proposed methods. Despite the noisy background and varying signatures in the GPR scans, the proposed method demonstrated promising results in object detection and segmentation. Computational results show that the improved MS R-CNN outperforms the other state-of-the-art methods.

Ground penetrating radar (GPR), as a non-invasive instrument, has been widely used in the civil field. The interpretation of GPR data plays a vital role in underground infrastructures to transfer raw data to the interested information, such as diameter. However, the diameter identification of objects in GPR B-scans is a tedious and labor-intensive task, which limits the further application in the field environment. The paper proposes a deep learning-based scheme to solve the issue. First, an adaptive target region detection (ATRD) algorithm is proposed to extract the regions from B-scans that contain hyperbolic signatures. Then, a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) framework is developed that integrates Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network to extract hyperbola region features. It transfers the task of diameter identification into a task of hyperbola region classification. Experimental results conducted on both simulated and field datasets demonstrate that the proposed scheme has a promising performance for diameter identification. The CNN-LSTM framework achieves an accuracy of 99.5% on simulated datasets and 92.5% on field datasets.

Ground-penetrating radar (GPR), as a non-invasive instrument, has been widely used in civil engineering. In GPR B-scan images, there may exist random noise due to the influence of the environment and equipment hardware, which complicates the interpretability of the useful information. Many methods have been proposed to eliminate or suppress the random noise. However, the existing methods have an unsatisfactory denoising effect when the image is severely contaminated by random noise. This paper proposes a multi-scale convolutional autoencoder (MCAE) to denoise GPR data. At the same time, to solve the problem of training dataset insufficiency, we designed the data augmentation strategy, Wasserstein generative adversarial network (WGAN), to increase the training dataset of MCAE. Experimental results conducted on both simulated, generated, and field datasets demonstrated that the proposed scheme has promising performance for image denoising. In terms of three indexes: the peak signal-to-noise ratio (PSNR), the time cost, and the structural similarity index (SSIM), the proposed scheme can achieve better performance of random noise suppression compared with the state-of-the-art competing methods (e.g., CAE, BM3D, WNNM).

Ground Penetrating Radar (GPR) is a non-destructive detection technology based on electromagnetic waves. However, the presence of clutter during the data acquisition process has a serious impact on the quality of GPR B-scan images. In particular, when strong clutter such as rebar mesh is present in the surface layer of the road area being scanned, it can be challenging to locate and identify road defects such as voids and faults in the acquired images. To suppress the surface rebar clutter and recover the disturbed defect echoes, we designed an unsupervised SuppRebar-GAN network. Meanwhile, to verify the accuracy of the location and type of its recovered echoes, we also set up an EC-Yolov7 network to localize and identify the recovered echoes and generate corresponding heatmaps based on the detection results to estimate the true location of the roadway defect in the subsurface. In this study, simulated and measured data of four typical road defects, namely void, fault, uncompact, and crack, were collected for training and testing the two networks. The experimental results show that both networks can perform their respective functions excellently without retraining in the face of test datasets collected in other simulations or real test scenarios that are very different from the initial training set.

Gesture recognition that is based on high-resolution radar has progressively developed in human-computer interaction field. In a radar recognition-based system, it is challenging to recognize various gesture types because of the lacking of gesture transversal feature. In this paper, we propose an integrated gesture recognition system that is based on frequency modulated continuous wave MIMO radar combined with deep learning network for gesture recognition. First, a pre-processing algorithm, which consists of the windowed fast Fourier transform and the intermediate-frequency signal band-pass-filter (IF-BPF), is applied to obtain improved Range Doppler Map. A range FFT based MUSIC (RFBM) two-dimensional (2D) joint super-resolution estimation algorithm is proposed to obtain a Range Azimuth Map to obtain gesture transversal feature. Range Doppler Map and Range Azimuth Map then respectively form a Range Doppler Map Time Sequence (RDMTS) and a Range Azimuth Map Time Sequence (RAMTS) in gesture recording duration. Finally, a Dual stream three-dimensional (3D) Convolution Neural Network combined with Long Short Term Memory (DS-3DCNN-LSTM) network is designed to extract and fuse features from both RDMTS and RAMTS, and then classify gestures with radial and transversal change. The experimental results show that the proposed system could distinguish 10 types of gestures containing transversal and radial motions with an average accuracy of 97.66%.

Saliency map generation in synthetic aperture radar (SAR) imagery has become a promising research area, since it has a close relationship with quick potential target identification, rescue services, etc. Due to the multiplicative speckle noise and complex backscattering in SAR imagery, producing satisfying results is still challenging. This paper proposes a new saliency map generation approach for SAR imagery using Bayes theory and a heterogeneous clutter model, i.e., the $G_0$ model. With Bayes theory, the ratio of the probability density functions (PDFs) in the target and background areas contributes to the saliency. Local and global background areas lead to different saliency measures, i.e., local saliency and global saliency, which are combined to make a final saliency measure. To measure the saliency of targets of different sizes, multiscale saliency enhancement is conducted with different region sizes of target and background areas. After collecting all of the salient regions in the image, the result is refined by considering the image’s immediate context. The saliency of regions that are far away from the focus of attention is suppressed. Experimental results with two single-polarization and two multi-polarization SAR images demonstrate that the proposed method has better speckle noise robustness, higher accuracy, and more stability in saliency map generation both with and without the complex background than state-of-the-art methods. The saliency map accuracy can achieve above 95% with four datasets, which is about 5–20% higher than other methods.