Detection of objects with diverse geometric shapes in GPR images using deep-learning methods

Orhan Apaydın; Turgay İşseven

doi:10.1515/geo-2022-0685

Article Open Access

Detection of objects with diverse geometric shapes in GPR images using deep-learning methods

Orhan Apaydın and Turgay İşseven

Published/Copyright: October 10, 2024

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Open Geosciences Volume 16 Issue 1

Abstract

Buried objects with regular geometric shapes appear as hyperbolic structures in ground-penetrating radar (GPR) images. The shapes of these hyperbolic structures differ depending on the geometric shapes of the objects. In this study, current deep learning-based object detection algorithms such as Faster R-CNN, YOLOv5, and single-shot detector are used to detect hyperbolic structures in GPR images and classify the buried object according to its geometric shape. A mixed data set is produced for training the models. A GPR measurement device is designed with a vector network analyzer and Vivaldi antenna pair to be used in laboratory measurements. Objects with rectangular and cylindrical geometric shapes are placed under the table and measurements are performed. The measurement scenarios created in the laboratory are modeled in the gprMax program and synthetic GPR data are produced. Data augmentation techniques such as flipping and resizing are applied to expand the data set. As a result of the training, three models successfully detect the objects and classify them according to their geometric shapes. The Faster R-CNN model gives the most accurate detection and classification with the metrics classification loss = 5.4 × 10⁻³, localization loss = 9 × 10⁻³, regularization loss = 5.1 × 10⁻⁵, mAP@0.5 = 1, and mAP@0.5:0.95 = 1.

Keywords: deep learning; ground-penetrating radar; object detection; rectangular and cylindrical objects; object classification

1 Introduction

Ground-penetrating radar (GPR) is a non-destructive geophysical exploration method used in near-surface investigations [1,2]. It is frequently used for detecting and positioning objects buried in the subsurface, such as mines, water pipes, and graves [3,4]. Object detection and classification in GPR images can be implemented using deep learning-based methods [5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23]. Three of these methods are Faster R-CNN, You Only Look Ones (YOLOv5), and single-shot detector (SSD). These techniques are capable of detecting multiple objects in an image and accurately presenting them within bounding boxes. Near-surface buried objects appear as hyperbola in GPR images. Rectangular and cylindrical-shaped objects create variations in the hyperbolic structures within GPR images. In this study, we investigate the detection and classification of objects with diverse geometric shapes in GPR images. The dataset includes radargrams obtained from real measurements collected by a radar system consisting of a vector network analyzer (VNA) and Vivaldi antenna pair in a laboratory environment, as well as synthetic radargrams created in the gprMax program. Rectangular-shaped aluminum box and cylindrical-shaped iron bar areplaced beneath the table, and GPR measurements are performed. The hyperbolic structures of these objects are detected in GPR images, and it is found that the object is either rectangular or cylindrical. Thus, the studies in the literature have been extended in terms of detecting different geometric-shaped structures in GPR images.

2 Methods

2.1 Faster R-CNN

Most of the object detection algorithms are based on the CNN method [24]. CNN is one of the most popular feature extraction methods according to its remarkable successful results. In a simple CNN structure, it includes a convolution layer, a pooling layer, and a flattening layer [25]. Region-based CNN (R-CNN) is introduced by Girshick et al. [26] to get better results in object detection. In the first step of R-CNN, object proposals are produced by using selective search (SS). In the second step, possible object areas within each proposal are transformed into a feature vector and then passed to a classifier and also to a regressor. Since these operations are performed for each proposal, it causes an intense computational cost [7]. To overcome this problem, SSPnet and Fast R-CNN recommend that all input images pass through CNN once. Ren et al. [27] introduced the Faster R-CNN object detection method, which is based on R-CNN. Faster R-CNN uses the region proposal network (RPN) instead of SS. RPN makes the Faster R-CNN’s performance better than models that use SS because of the most accurate results and fast algorithm speed [11]. Faster R-CNN consists of two main parts (Figure 1): (1) an RPN that extracts information about the possible locations of objects and (2) the Fast R-CNN network where the object is classified and bounding boxes are found. These two parts work as a whole by sharing their parameters. In the classification part, it will output that if the confidence score is greater than 0.7, it is an object, and if it is less than this value, it is not an object.

Figure 1

Network architecture of Faster R-CNN. There are two learnable stages: region proposal network and classification network.

In this study, the Faster R-CNN method is applied using TensorFlow Object detection API. Parameters used in network training are tuned as Epoch = 10,000, learning tare = 1 × 10⁻², and batch size = 3.

2.2 YOLOv5

Another network mostly used for object detection is YOLO (You Only Look Once) [28]. YOLO is a CNN-based method. YOLOv5 is the fifth version of YOLO [29]. Multiple objects in an image can be detected by using YOLOv5. This method works with faster training and inference time than other detection networks like SSD and Faster R-CNN and makes it more suitable for real-time applications [30]. However, in addition to its fast work, its accuracy rates in detecting small-size objects including handicaps when compared to other object detection methods.

The main idea of the method is to use the whole picture as input data and pass it through the network to generate bounding boxes containing five predictions: x, y, w, h, and confidence values. While the (x, y) coordinates from the estimates describe the center of the boxes according to the grid cell, the width and heightarerelative values to the whole image [28,30,31].

In the method, the image is divided into S × S grid cells and CNN is performed in each region. A confidence score is issued indicating whether the regions contain objects. Generally, it is best if the confidence score is equal to the point of IOU between the predicted box and the ground truth. Each grid cell contains B bounding boxes which have a class probability (C). All predictions are encoded as an S × S × (B × 5 × C) size tensor [28].

The YOLOv5 consists of three main blocks (Figure 2): (1) “Backbone,” part of feature extraction with CNN, (2) “Neck,” at this stage aggregation of image features is performed, and (3) "Head" section, the object detection predictions are made and the results are shown from these predictions [5].

Figure 2

Network architecture of YOLOv5.

Hyperparameters used in YOLOv5 network training are tuned as Epoch = 210, learning rate = 1 × 10⁻², and batch size = 4.

2.3 SSD

Unlike traditional methods that require multiple passes through an image, SSD performs object detection in a single shot, thanks to its innovative architecture (Figure 3). By utilizing a combination of 1 × 1 and 3 × 3 filters, SSD generates feature maps essential for detecting objects of various sizes and shapes [32,33]. Furthermore, the SSD algorithm incorporates the MultiBox technique, which involves utilizing predefined bounding boxes placed over actual regions within an image. By segmenting the image into grids of varying sizes, SSD effectively detects both large and small objects, enhancing its versatility. Each predicted bounding box is assigned a confidence score for label prediction, with the highest-scoring box being selected.

Figure 3

SSD model architecture.

This process is facilitated by multi-scale feature maps, where convolutional layers of different cell sizes are employed to detect objects across various scales. To train the SSD model, the TensorFlow Object Detection API with the Resnet50 base model is utilized. The training parameters are carefully configured, with a batch size of 3, a learning rate of 1 × 10⁻⁴, and 3,000 epochs.

3 Dataset

In this section, the dataset used in object detection algorithms Faster R-CNN, YOLOv5, and SSD is explained. Dataset has a major role in object detection models. The dataset used in this study consists of measurements performed in the laboratory environment and simulated data created by using gprMax. To increase the variety of the dataset and the accuracy of the algorithms, data augmentation is performed with flipping and image resizing methods. The dataset information used in similar object detection studies in the literature is given in Table 1.

Table 1

Dataset information in the literature and this study

GPR object detection studies	Total	gprMax	Lab and field	Train	Valid.	Test
Lei et al. [7]	5,866	5,236	630	—	—	—
Kumlu [10]	55	55	—	40	10	5
Dewantara and Parnadi [12]	80	78	2	72	—	8
Kafedziski et al. [17]	157	—	—	—	—	—
Pham and Lefèvre [18]	150	50	100	—	—	—
Cui et al. [9]	3,000	—	—	2,100	600	300
Xiao et al. [11]	4,688	—	—	3,281	—	1,407
Gong and Zhang [8]	200	200	—	160	—	40
In this study	357	300	57	249	72	36

3.1 Laboratory measurements

In the laboratory environment, measurements are performed with a GPR system using a Copper Mountain TR/1,300 brand and model VNA and Vivaldi antenna pair that can operate at the appropriate frequency for measurements. In measurements, cylindrical and rectangular objects are placed beneath the table at different depths (Figure 4).

Figure 4

Laboratory measurement setup: (a) rectangular shape and (b) cylindrical shape object.

The frequency band of the VNA device used for the measurements is 600–1,300 MHz, the center frequency is 950 MHz, the number of data points in an A-scan is 1,000, the measurement intervals are 2 cm, the total measurement length is 120 cm, and the GPR data length is 20 ns. The depth of the targets from the table varies between 5 and 30 cm. A total of 19 GPR data are collected for 10 rectangular objects (boxes) and 9 cylindrical objects (iron pipes). Attention has been given to enhancing the electromagnetic reflectivity of objects by using metallic materials. To make the hyperbolas of the box more visible in GPR images, it is covered with aluminum foil with high electromagnetic reflectivity. As seen in Figure 5, the hyperbolas formed by the box and the iron bar in the GPR images exhibit distinct differences.

Figure 5

GPR images of (a) rectangular and (b) cylindrical objects.

Due to the flat surface on the top of the rectangular-shaped box, the apex of the hyperbola it forms is also planar. In contrast, the interaction area on the cylindrical pipe is much narrower, and as a result, this flat area is not visible at the top of its hyperbola.

3.2 Synthetic data

To expand the dataset, synthetic GPR images are generated through simulation. The gprMax program, capable of simulating electromagnetic waves by solving Maxwell’s equations using the finite difference time domain method [34,35], is used to generate synthetic data. It is essential to emphasize that the inclusion of additional simulated data can lead to more accurate and reliable detection results, marked by an increase in successful detections and a reduction in false alarms. This observation supports the scheme’s intended purpose and affirms its effectiveness [18].

The Vivaldi antennas used in the measurements are modeled in the gprMax program with the same size and material properties as the reality and used in the simulation to obtain the closest results to the data obtained in laboratory measurements. The table, aluminum-covered box, and iron bar used in laboratory measurements are modeled similarly and used in the simulation (Figure 6).

Figure 6

Simulation setup in gprMax: (a) rectangular shape aluminum covered box and (b) cylindrical shape iron pipe.

The center frequency (950 MHz), measurement interval (2 cm), measurement line length (120 cm), and depth values of the objects (5–30 cm) used in the simulation are the same as the values used in laboratory measurements. A total of 100 synthetic GPR images are created. The pixel values of GPR images are 250 × 1,000. An example of a simulation GPR image of the box and the cylinder is shown in Figure 7.

Figure 7

Synthetic GPR images of (a) rectangular and (b) cylindrical objects.

Similar to the laboratory measurements, the hyperbolic shape anomalies of the aluminum-covered box and the iron bar are obtained differently from each other. Additionally, the waveforms of hyperbolas resemble the waveforms obtained in the laboratory measurements.

3.3 Data augmentation

Data augmentation is generally applied in object detection applications to increase detection accuracy, with resizing and flipping being the most well-known and commonly used techniques. The input image of 250 × 1,000 is set to 300 × 1,000 by using resizing. Figure 8 shows a GPR image that has been flipped and resized.

Figure 8

(a) Original GPR image, (b) flipped, and (c) resized GPR images.

As a result of augmentation, the total number of laboratory and simulation GPR data increased from 119 to 357 images. The dataset is distributed as 70% for training, 20% for validation, and 10% for testing. The dataset includes two classes, one of which is ‘cylindrical’ with 183 instances, and the other one is ‘rectangular’ with 174 instances.

4 Results and discussion

The Faster R-CNN application is configured with epoch = 10,000, batch size = 3, learning rate = 1 × 10⁻², and the training took 68 min. At the end of the training, the mAP@0.5 and mAP@0.75 values are obtained as 1, and the training classification loss, localization loss, and regularization loss values are obtained as 5.4 × 10⁻³, 9 × 10⁻³, and 5.1 × 10⁻⁵ respectively. We tested the trained model with confidence >0.7. As can be seen in Figure 9(a)–(c), the values giving the lowest loss values are selected in the hyperparameter test performed to select the training parameters of the Faster RCNN application.

Figure 9

Results of the hyperparameter test to determine the parameters to be used in Faster RCNN, YOLOv5, and SSD training. The tests are carried out using four different batch sizes and three different learning rates. (a, d, and g) Classification losses, (b, e, and h) localization losses, (c and i) regularization losses, and (f) objectness loss results.”

In the YOLOv5 application, the epoch is set to 210, batch size to 4, and learning rate to 1 × 10⁻⁵. The training process took 9.84 min in total. At the end of the YOLOv5 training, the mAP@0.5 and mAP@0.5:0.95 values are obtained as 0.995 and 0.779, respectively, and the classification loss, localization loss, and objectness loss values are obtained as 1.9 × 10⁻³, 1.5 × 10⁻², and 7.2 × 10⁻³, respectively.

The third network, SSD, is trained with these tuned configuration parameters: epoch = 3,000, batch size = 3, learning rate = 1 × 10⁻⁴, and the training took 17 min. The mAP@0.5 and mAP@0.75 values are obtained as 1 and 0.79, respectively. The training classification loss, localization loss, and regularization loss values are obtained as 2.1 × 10⁻¹, 9.4 × 10⁻³, and 1.9 × 10⁻¹, respectively. The regularization loss of the SSD method (Figure 10(i)) decreased in a different trend compared to the graphs of other models due to the simple structure of the model. Although a similar pattern of regularization loss decrease was observed in the [36] study, the detection results were successful.

Figure 10

Train and validation loss values obtained in each epoch as a result of Faster RCNN, YOLOv5, and SSD training. (a, d, and g) Localization losses, (b, e, and h) classification losses, (c and i) regularization losses, and (f) objectness loss results.

For all three methods, the training and validation loss values decreased smoothly at each epoch throughout the training process (Figure 10). Recall, precision, and F1-score metrics and detection times are given in Table 2.

Table 2

The performance metrics and detection times of Faster R-CNN, YOLOv5, and SSD

Networks	Geometrical shape	Precision (%)	Recall (%)	F1 score (%)	Detection time (s)
Faster R-CNN	Rectangular	100.0	100.0	100.0	2.2
	Cylindrical	100.0	100.0	100.0
YOLOv5	Rectangular	70.6	100.0	82.8	0.0028
	Cylindrical	82.6	94.8	88.3
SSD	Rectangular	74.1	58.9	65.6	1.9
	Cylindrical	62.7	77.1	69.2

In this study, hyperbolas are detected and classified based on the geometric shapes of the objects in GPR images with high accuracy rates. Both Faster R-CNN and YOLOv5 give successful results for detection and classification, but SSD has bad performance scores (Figure 11). YOLOv5 is more suitable for real-time detection applications with 2.8 ms detection time. Faster R-CNN, with its high accuracy detection performance, can be used to detect objects in the radargrams obtained after the GPR applications.

Figure 11

Some object detection and classification test results in GPR images of (a) Faster R-CNN, (b) YOLOv5, and (c) SSD models. Models output detected and classified objects with a bounding box and confidence score in percentage.

In studies [6,7,10,11,12,13,14,15,16], hyperbolas of objects in GPR images are detected and not classified by using only one of the Faster R-CNN and YOLOv5 methods [8]. It successfully detected objects with different geometric shapes, but the model used in the training is Faster R-CNN and the dataset does not consist of laboratory or field data, but only synthetic data. In this study, Faster R-CNN, YOLOv5, and SSD methods are tested separately to find out which method gave better results for detection and classification.

Acknowledgments

We would like to thank Prof. Dr. Selçuk Paker for his ideas and support in laboratory measurements and Prof. Dr. Işın Erer for providing the GPR device.

Funding information: Authors state no funding involved.
Author contributions: Orhan Apaydın: writing – original draft preparation, software visualization, data curation, formal analysis, methodology. Turgay Işseven: conceptualization, methodology, project administration, validation.
Conflict of interest: Authors state no conflict of interest.
Data availability statement: The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

References

[1] Jol HM. Ground penetrating radar theory and applications. Amsterdam, Netherlands: Elsevier; 2008.Search in Google Scholar

[2] Daniels DJ. Ground penetrating radar. London, UK: IET; 2004.10.1049/PBRA015ESearch in Google Scholar

[3] Ho K, Collins LM, Huettel LG, Gader PD. Discrimination mode processing for EMI and GPR sensors for hand-held land mine detection. IEEE Trans Geosci Remote Sens. 2004;42:249–63.10.1109/TGRS.2003.817804Search in Google Scholar

[4] Pettinelli E, Di Matteo A, Mattei E, Crocco L, Soldovieri F, Redman JD, et al. GPR response from buried pipes: Measurement on field site and tomographic reconstructions. IEEE Trans Geosci Remote Sens. 2009;47:2639–45.10.1109/TGRS.2009.2018301Search in Google Scholar

[5] Qiu Z, Zhao Z, Chen S, Zeng J, Huang Y, Xiang B. Application of an improved YOLOv5 algorithm in real-time detection of foreign objects by ground penetrating radar. Remote Sens. 2022;14:1895.10.3390/rs14081895Search in Google Scholar

[6] Zhao Y, Zhang J, Li H, Gu X, Li Z, Zhang S. Enhanced automatic root recognition and localization in GPR images through a YOLOv4-based deep learning approach. Med Biol Eng Comput. 2022;60:1–14.10.1109/TGRS.2022.3181202Search in Google Scholar

[7] Lei W, Hou F, Xi J, Tan Q, Xu M, Jiang X, et al. Automatic hyperbola detection and fitting in GPR B-scan image. Autom Constr. 2019;106:102839.10.1016/j.autcon.2019.102839Search in Google Scholar

[8] Gong Z, Zhang H. Research on GPR image recognition based on deep learning. MATEC Web of Conferences. EDP Sciences; 2020. p. 03027.10.1051/matecconf/202030903027Search in Google Scholar

[9] Cui F, Ning M, Shen J, Shu X. Automatic recognition and tracking of highway layer-interface using Faster R-CNN. J Appl Geophys. 2022;196:104477.10.1016/j.jappgeo.2021.104477Search in Google Scholar

[10] Kumlu D. GPR image recovery effect on faster R-CNN-based buried target detection. J Electromagn Eng Sci. 2022;22:591–8.10.26866/jees.2022.5.r.127Search in Google Scholar

[11] Xiao J, Gao Q, Ling Y, Yan J, Liu B. Research on hyperbola detection and fitting in GPR B-scan image. 2021 International Conference on Communications, Information System and Computer Engineering (CISCE). IEEE; 2021. p. 266–70.10.1109/CISCE52179.2021.9445915Search in Google Scholar

[12] Dewantara D, Parnadi WW. Automatic hyperbola detection and apex extraction using convolutional neural network on GPR data. Journal of Physics: Conference Series. IOP Publishing; 2022. p. 012027.10.1088/1742-6596/2243/1/012027Search in Google Scholar

[13] Liu H, Yue Y, Liu C, Spencer Jr B, Cui J. Automatic recognition and localization of underground pipelines in GPR B-scans using a deep learning model. Tunn Undergr Space Technol. 2023;134:104861.10.1016/j.tust.2022.104861Search in Google Scholar

[14] Li S, Gu X, Xu X, Xu D, Zhang T, Liu Z, et al. Detection of concealed cracks from ground penetrating radar images based on deep learning algorithm. Constr Build Mater. 2021;273:121949.10.1016/j.conbuildmat.2020.121949Search in Google Scholar

[15] Li X, Liu H, Zhou F, Chen Z, Giannakis I, Slob E. Deep learning–based nondestructive evaluation of reinforcement bars using ground‐penetrating radar and electromagnetic induction data. Comput Civ Infrastruct Eng. 2021;37:1834–53.10.1111/mice.12798Search in Google Scholar

[16] Li Y, Zhao Z, Luo Y, Qiu Z. Real-time pattern-recognition of GPR images with YOLO v3 implemented by tensorflow. Sensors. 2020;20:6476.10.3390/s20226476Search in Google Scholar PubMed PubMed Central

[17] Kafedziski V, Pecov S, Tanevski D. Detection and classification of land mines from ground penetrating radar data using faster R-CNN. 2018 26th Telecommunications Forum (TELFOR). IEEE; 2018. p. 1–4.10.1109/TELFOR.2018.8612117Search in Google Scholar

[18] Pham M-T, Lefèvre S. Buried object detection from B-scan ground penetrating radar data using Faster-RCNN. IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium. IEEE; 2018. p. 6804–7.10.1109/IGARSS.2018.8517683Search in Google Scholar

[19] Huang J, Yang X, Zhou F, Li X, Zhou B, Lu S, et al. A deep learning framework based on improved self‐ supervised learning for ground‐penetrating radar tunnel lining inspection. Comput‐Aided Civ Infrastruct Eng. 2023;39:814–33.10.1111/mice.13042Search in Google Scholar

[20] Liu B, Zhang J, Lei M, Yang S, Wang Z. Simultaneous tunnel defects and lining thickness identification based on multi-tasks deep neural network from ground penetrating radar images. Autom Constr. 2023;145:104633.10.1016/j.autcon.2022.104633Search in Google Scholar

[21] Qin H, Zhang D, Tang Y, Wang Y. Automatic recognition of tunnel lining elements from GPR images using deep convolutional networks with data augmentation. Autom Constr. 2021;130:103830.10.1016/j.autcon.2021.103830Search in Google Scholar

[22] Yue Y, Liu H, Lin C, Meng X, Liu C, Zhang X, et al. Automatic recognition of defects behind railway tunnel linings in GPR images using transfer learning. Measurement. 2024;224:113903.10.1016/j.measurement.2023.113903Search in Google Scholar

[23] Yue Y, Liu H, Meng X, Li Y, Du Y. Generation of high-precision ground penetrating radar images using improved least square generative adversarial networks. Remote Sens. 2021;13:4590.10.3390/rs13224590Search in Google Scholar

[24] Li K, Ma W, Sajid U, Wu Y, Wang G. Object detection with convolutional neural networks. Deep learning in computer vision. Elsevier; 2020. p. 41–62.10.1201/9781351003827-2Search in Google Scholar

[25] O'Shea K, Nash RJ. An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458; 2015.Search in Google Scholar

[26] Girshick R, Donahue J, Darrell T, Malik J. Rich feature hierarchies for accurate object detection and semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2014; p. 580–7.10.1109/CVPR.2014.81Search in Google Scholar

[27] Ren S, He K, Girshick R, Sun J. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence. 2015;28:1137–49.10.1109/TPAMI.2016.2577031Search in Google Scholar PubMed

[28] Redmon J, Divvala S, Girshick R, Farhadi A. You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2016. p. 779–88.10.1109/CVPR.2016.91Search in Google Scholar

[29] Jocher G, Stoken A, Borovec J, Changyu L, Hogan A, Diaconu L, et al. ultralytics/yolov5: v3. 1-bug fixes and performance improvements. Zenodo; 2020.Search in Google Scholar

[30] Yao J, Qi J, Zhang J, Shao H, Yang J, Li X. A real-time detection algorithm for Kiwifruit defects based on YOLOv5. Electronics. 2021;10:1711.10.3390/electronics10141711Search in Google Scholar

[31] Bochkovskiy A, Wang C-Y, Liao H-Y. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934; 2020.Search in Google Scholar

[32] Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, et al. Ssd: Single shot multibox detector. Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer; 2016. p. 21–37.10.1007/978-3-319-46448-0_2Search in Google Scholar

[33] Sharma A, Mishra T, Kukade J, Golwalkar A, Tomar H. Object detection using tensorflow. International Conference on ICT for Sustainable Development. Springer; 2023. p. 343–52.10.1007/978-981-99-6568-7_31Search in Google Scholar

[34] Warren C, Giannopoulos A, Giannakis I. gprMax: Open source software to simulate electromagnetic wave propagation for Ground Penetrating Radar. Comput Phys Commun. 2016;209:163–70.10.1016/j.cpc.2016.08.020Search in Google Scholar

[35] Giannopoulos A. Modelling ground penetrating radar by GprMax. Constr Build Mater. 2005;19:755–62.10.1016/j.conbuildmat.2005.06.007Search in Google Scholar

[36] Umair M, Foo Y-L. Industrial safety helmet detection using single shot detectors models and transfer learning. Multimedia University Engineering Conference (MECON 2022). Atlantis Press; 2022. p. 390–400.10.2991/978-94-6463-082-4_34Search in Google Scholar

Received: 2023-11-10

Revised: 2024-04-29

Accepted: 2024-06-03

Published Online: 2024-10-10

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/geo-2022-0685

Keywords for this article

deep learning; ground-penetrating radar; object detection; rectangular and cylindrical objects; object classification

Creative Commons

BY 4.0