1 Improving the Anomaly Detection in GPR Images by Fine-Tuning CNNs with Synthetic Data

2025-04-30 0 0 5.35MB 11 页 10玖币

侵权投诉

Improving the Anomaly Detection in GPR Images

by Fine-Tuning CNNs with Synthetic Data

Xiren Zhou, Shikang Liu, Ao Chen, Yizhan Fan, Huanhuan Chen, Senior Member, IEEE

Abstract—Ground Penetrating Radar (GPR) has been widely

used to estimate the healthy operation of some urban roads and

underground facilities. When identifying subsurface anomalies

by GPR in an area, the obtained data could be unbalanced,

and the numbers and types of possible underground anomalies

could not be acknowledged in advance. In this paper, a novel

method is proposed to improve the subsurface anomaly detection

from GPR B-scan images. A normal (i.e. without subsurface

objects) GPR image section is ﬁrstly collected in the detected area.

Concerning that the GPR image is essentially the representation

of electromagnetic (EM) wave and propagation time, and to

preserve both the subsurface background and objects’ details, the

normal GPR image is segmented and then fused with simulated

GPR images that contain different kinds of objects to generate

the synthetic data for the detection area based on the wavelet

decompositions. Pre-trained CNNs could then be ﬁne-tuned with

the synthetic data, and utilized to extract features of segmented

GPR images subsequently obtained in the detection area. The

extracted features could be classiﬁed by the one-class learning

algorithm in the feature space without pre-set anomaly types

or numbers. The conducted experiments demonstrate that ﬁne-

tuning the pre-trained CNN with the proposed synthetic data

could effectively improve the feature extraction of the network

for the objects in the detection area. Besides, the proposed method

requires only a section of normal data that could be easily

obtained in the detection area, and could also meet the timeliness

requirements in practical applications.

Index Terms—Ground Penetrating Radar, B-scan Image,

Anomaly Detection, Buried Asset Detection, Data Processing.

I. INTRODUCTION

Ground Penetrating Radar (GPR) has become increasingly

important as a nondestructive tool to estimate the healthy

operation of some urban roads and underground facilities,

which makes the use of the transmission and reﬂection of

the electromagnetic (EM) waves to detect dielectric properties

changes in host materials [1]. By arranging received EM

waves horizontally in a temporal or spatial relationship, and

representing the wave intensities with corresponding grayscale

values, a GPR B-scan image could be obtained [2] as shown

in Fig. 1. The subsurface situation or existing objects could be

estimated by interpreting different shapes or characteristics on

the obtained GPR B-scan image [3]. But due to the refraction

and reﬂection of EM waves, the various shapes on the GPR

image1are not actually the same as the actual objects. Besides,

Xiren Zhou, Shikang Liu, Ao Chen, Yizhan Fan, and Huanhuan

Chen are with School of Computer Science and Technology, University

of Science and Technology of China, Hefei, 230027, China; e-

mail: zhou0612@mail.ustc.edu.cn, qq1321401109@mail.ustc.edu.cn,

chenao57@mail.ustc.edu.cn, fyz666@mail.ustc.edu.cn, hchen@ustc.edu.cn.

Corresponding Author: Huanhuan Chen

1The GPR image in this paper refers to the GPR B-scan image as shown

in Fig. 1(c).

the system noise, the heterogeneity of the medium, and mutual

wave interactions also make it challenging to automatically

cope with the GPR images obtained in the detection area [4].

The host of the

utilized GPR

The utilized GPR antenna

The moved direction of

the utilized GPR antenna

(a)

The host of the

utilized GPR

The utilized GPR antenna

The moved direction of

the utilized GPR antenna

(b)

(c)

Fig. 1. (a) and (b) present a real detection along an urban road using GPR.

(a) shows the host and antenna of the utilized GSSI GPR. (b) shows the

actual detection scene of a road. The GPR antenna is moved along the road

(the red arrow) to obtain the GPR B-scan image, and the collected image is

transformed to the host. (c) illustrates the generation of a GPR B-can image.

The abscissa in the left B-scan image is the detection position, and the ordinate

is the time from emission to reception of the electromagnetic wave. The right

side is a received wave in the left B-scan image. As shown in the bottom

of the right side, different gray values correspond to different magnitudes

of the EM wave. The gray value of each pixel represents the corresponding

amplitude at this position and time.

The interpretation of GPR B-scan images could be roughly

grouped into two categories: extracting and ﬁtting hyper-

bolic characteristics on B-scan images, and identifying non-

hyperbolic shapes. The hyperbolic characteristics are gener-

ated by underground objects with circular cross sections (e.g.

tree roots, pipes, etc.). In addition to hyperbolic characteristics,

non-hyperbolic shapes could be more common targets in GPR

B-scan images when imaging the subsurface. These shapes

could be formed by different kinds of subsurface media or

objects, including subsurface cavities, moisture damage, loose

media, etc. Different subsurface media or objects would gen-

erate shapes with different characteristics on the GPR image.

Even for the same type of underground objects, the generated

shapes on the GPR image vary in different underground

environments or by different GPR [5].

To identify non-hyperbolic shapes in GPR images, signal or

image processing methods [6], [7], and Convolutional Neural

Network (CNN) [8]–[10] are utilized. But there are still some

arXiv:2210.11833v2 [cs.CV] 22 Nov 2022

One-class learning in the

feature space

Fine-tuning Extracting

feature

Pre-trained Convolutional

Neural Network (CNN)

Pre-trained Convolutional

Neural Network (CNN)

The GPR images subsequently obtained in the detected area

is swiped with a sliding window

The fine-tuned CNN is utilized to extract

features of the GPR image in each window

Image

fusing

A section of normal GPR image of the detected area

Simulated GPR B-scan images with

various underground structures

Synthesized Data

Segmented

Extracting

feature

Pre-trained Convolutional

Neural Network (CNN)

Pre-trained Convolutional

Neural Network (CNN)

The GPR images subsequently obtained in the

detected area is swiped with a sliding window

The fine-tuned CNN is utilized to extract

features of the GPR image in each window

Image

fusing

A section of normal GPR image of the detected area

Synthesized Data

Fine-tuning

Segmented

One-class learning in the

feature space

Simulated GPR B-scan images with

various underground structures

Fig. 2. This ﬁgure shows the procedure of the proposed method, which could be divided into four steps. 1) A section of a normal GPR image without any

subsurface objects is obtained in the detection area. 2) The normal GPR image section is segmented, and fused with simulated GPR images containing various

kinds of subsurface objects to generate synthesized data for the detection area. 3) The pre-trained CNN is ﬁne-tuned with the generated synthesized data.

The subsequently obtained GPR image is swiped with a sliding window, and the ﬁne-tuned CNN is utilized to extract features of the GPR image in each

window. 4) The extracted features are ﬁnally classiﬁed by the one-class learning algorithm in the feature space, and the corresponding underground anomalies

are detected.

practical issues that need to be considered when identifying

underground anomalies from GPR images obtained in an

unknown area: 1) The obtained data could be unbalanced,

since the amount of GPR image generated by subsurface

anomalies could be much smaller than the normal data, and

there could be even no abnormality in the detection area.

2) Only some normal GPR images without any underground

objects could be obtained at the beginning of the detection,

and it could be a common situation that the numbers and

types of the possible underground anomalies are rarely ac-

knowledged prior to fully analyzing the obtained GPR data.

3) The underground environments vary in areas, thus there is

no guarantee that training a model with data from an area or

datasets will improve its performance in other areas. 4) In real-

world applications, there could be timeliness requirements for

anomaly detection, and the locations of anomalies need to be

marked on-site for further repairs.

To address the above issues, a novel method based on CNN

and one-class learning is proposed in this paper to improve the

subsurface anomaly detection from the obtained GPR images.

The procedure of the proposed method is presented in Fig. 2,

which could be roughly divided into four steps: 1) A section

of a normal GPR image without any subsurface objects is

obtained in the detection area; 2) The normal GPR image

section is segmented, and fused with simulated GPR images

containing various kinds of subsurface objects to generate

synthesized data for the detection area; 3) The pre-trained

CNN is ﬁne-tuned with the generated data, and used to extract

features of the subsequently obtained GPR images; 4) The one-

class learning algorithm is utilized to identify anomalies from

the extracted features.

Existing studies [11]–[14] on transfer learning have shown

that proper transfer of networks could effectively improve the

learning results in the case of insufﬁcient training data, and any

pre-trained CNN could be adopted as a feature extractor for

further feature learning [15]–[17]. The features extracted from

different inputs constitute a feature space, and the distance

between two features in the feature space reﬂects the difference

between the two corresponding inputs. To improve the CNN’s

ability to identify different types of inputs in a speciﬁc task,

that is, to increase the inter-class distance in the feature space,

the pre-trained CNN could be further trained or ﬁne-tuned

with the data that contains both the background and target

details of this task [18], [19]. Pre-trained CNNs with adequate

ﬁne-tuning could perform at least as well as full training, and

the dependence on the amount of data could be signiﬁcantly

reduced [20]. However, as aforementioned, when detecting an

area, only a section of a normal GPR image could be obtained,

which is not enough to provide the type and amount of data

for ﬁne-tuning. Unlike visual images, GPR images have their

own physical meaning. The conducted experiments one real-

world dataset show limited effects of the feature extraction of

GPR images using CNNs only trained on visual images.

Nonetheless, by utilizing appropriate image fusion tech-

niques, remote sensing images that are very close to real

ones could be synthesized, and these generated data could

be used for change detection and object recognition [21]–

[23]. The GPR simulation software such as GprMax [24]

could generate GPR images of various underground objects

in pre-set environments. Therefore, in this paper, a section of

a normal GPR image is ﬁrstly obtained in the detection area,

and then fused with some simulated GPR images containing

various kinds of underground objects to generate synthetic

data speciﬁcally designed for the subsurface environment of

this area. Concerning that the GPR image is essentially the

representation of EM wave intensity and propagation time

as Fig. 1(c), and to preserve both the subsurface background

and target details, the wavelet decompositions of the normal

and simulated images are merged to generate the synthetic

images that contain both the basic underground conditions

of the detection area and various characteristics generated by

underground objects. The conducted experiments demonstrate

that ﬁne-tuning pre-trained CNNs with the generated synthetic

data could improve the feature extraction of the network for

further learning, i.e. features extracted from different kinds of

objects have larger distances in the feature space.

Due to the unknown numbers and types of subsurface

anomalies that may arise, there should be no preset anomaly

types in real-world applications. In the proposed method, a

sliding window is constructed and swiped across the ob-

tained GPR image. The normal GPR B-scan images in the

window without any underground objects are mapped to the

feature space via the ﬁne-tuned CNN, and an initial one-

class classiﬁer is trained. For the subsequent features extracted

from the image in the sliding window, the trained classiﬁer

is continuously used for classiﬁcation. The abnormal data

is further trained by incremental one-class learning, where

more incremental classiﬁers could be obtained to classify the

features into classes. Thus the corresponding underground

object that generates the GPR image could be detected.

The main contributions of the proposed method could be

summarized as follows.

1) A section of a normal GPR image is obtained in the

detection area, segmented, and fused with simulated

GPR images to generate synthetic data that contains both

the basic underground conditions of the detection area

and various characteristics generated by underground

objects. The conducted experiments demonstrate that

ﬁne-tuning CNNs with the synthetic data could increase

the distance between features extracted from GPR im-

ages formed by different objects in the detection area.

2) Only a section of a normal GPR image of the detec-

tion area (without subsurface anomalies) is required to

perform the proposed method, instead of an amount of

normal and abnormal data that is difﬁcult to collect,

process, and label in real-world applications.

3) By performing one-class learning, there is not need to

know the number and types of subsurface anomalies that

may exist in the detection area in advance. The proposed

method could incrementally classify various types of

underground objects through the extracted features.

The rest of this paper is organized as follows. Some related

work about interpreting GPR images is presented in Section

II. The feature extraction is presented in Section III, including

generating synthetic images, and ﬁne-tuning a pre-trained

CNN. The one-class learning is introduced in Section IV.

Experiments are conducted and analyzed in Section V. Finally,

conclusions are drawn in Section VI.

II. RELATED WORK

Interpreting GPR B-scan images could be roughly grouped

into two categories: extracting and ﬁtting hyperbolic char-

acteristics on B-scan images, and identifying non-hyperbolic

shapes. The prevailing methodologies in hyperbolic recogni-

tion from GPR images include the Hough transform (HT)

[25]–[27], Machine Learning (ML) [28]–[30] and some meth-

ods that combine multiple approaches [31]–[35]. In our pre-

vious work [36], a GPR B-scan image interpreting model has

been proposed, which could estimate the radius and depth of

the buried pipelines by extracting and ﬁtting hyperbolic point

clusters from GPR B-scan images.

Besides hyperbolic characteristics, some existing studies

identify underground objects with non-hyperbolic shapes from

GPR data by signal processing or image recognition methods.

The frequency-domain-focusing (FDF) technology of synthetic

aperture radar (SAR) is utilized to aggregate scattered GPR

signals for acquiring testing images, where a low-pass ﬁlter

is designed to denoise primordial signals, and the proﬁles of

detecting objects are extracted via the edge detection technique

based on the background information [6]. Subsequently, a

formula is conducted to relate the hidden crack width with

the relative measured amplitude [7]. The use of this kind of

method generally requires knowledge of the basic conditions

of the underground medium in advance.

To locate and identify objects in GPR images, the Convolu-

tional Neural Network (CNN) is utilized in recent decades. Re-

garding the EM signals as an input value, CNN structures are

conducted to automatically localize several kinds of targets in

GPR data [9]. The You-Only-Look-Once (YOLO) [37] is also

utilized to detect potholes and crackings beneath the roads.

Zhang et al. propose a mixed deep CNN model combined

with the Resnet-50 base network and YOLO framework to

detect the moisture damage in GPR data. Subsequently, Liu et

al. propose a method combining the YOLO series with GPR

images to recognize the internal defects in asphalt pavement

[38]. When detecting the underground objects in a certain area,

it is difﬁcult to ensure that the existing training data of CNN

obtained from other areas or datasets is consistent with the

underground situation in this area or road. And only some

GPR images without any target objects could be obtained at

the beginning of the detection, resulting in insufﬁcient training

data of CNN. The Generative Adversarial Network (GAN)

[39] could be utilized to generate remote-sensing data, but the

main issue with this approach in our scenario is that training

the generative network for the detection area could be time-

consuming for on-site applications.

III. FEATURE EXTRACTION BY FINE-TUNED CNNS

In this section, the generation of synthetic data of the

detection area is ﬁrstly introduced. After that, the pre-trained

CNN is ﬁne-tuned with the synthetic and normal data to

enhance the feature extraction capabilities for the objects in

the detection area.

A. Generating the Synthetic Data for the detection area

1) The Data Sources: As aforementioned, the synthetic

data is fused from two sources: 1) The normal GPR image

obtained in the detection area without any underground ob-

jects; 2) The GPR images simulated by GprMax with various

kinds of buried objects.

When detecting an area (e.g. a pavement road), a GPR

image section without any buried objects could be easily

obtained. This image section could be used to describe the

basic subsurface environment of the detection area, and pro-

vide data for CNN to extract the features of the GPR image

in the area without underground objects. In the conducted

experiments of this paper, the GPR image section with a length

greater than 3000 pixels is collected, and more than 300 GPR

image segment with the horizontal length of 300 pixels is then

randomly selected in this GPR image section 2.

2Duplications could exist in the selected images.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1ImprovingtheAnomalyDetectioninGPRImagesbyFine-TuningCNNswithSyntheticDataXirenZhou,ShikangLiu,AoChen,YizhanFan,HuanhuanChen,SeniorMember,IEEEAbstractGroundPenetratingRadar(GPR)hasbeenwidelyusedtoestimatethehealthyoperationofsomeurbanroadsandundergroundfacilities.Whenidentifyingsubsurfaceanomaliesb...

展开>> 收起<<

1 Improving the Anomaly Detection in GPR Images by Fine-Tuning CNNs with Synthetic Data.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Improving the Anomaly Detection in GPR Images by Fine-Tuning CNNs with Synthetic Data

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: