A Hybrid Pattern Recognition System for Detecting Buried Object in GPR Images

Author Name(s): *U. Özkaya, L. Seyfi
Author Email: uozkaya@selcuk.edu.tr

Abstract

It is very difficult to evaluate the shape of the buried objects using Ground Penetrating Radar (GPR). In this study, 180 GPR B scan images were classified by using different classification algorithms. GPR-B scan images of objects with different shapes in various depths were obtained by using GprMax simulation program. Noise in these images were eliminated by Wavelet transform and then, Gray Level Co-occurrence Matrices (GLCM), run length matrix (RLM) and Autocorrelation function (ACF) features were extracted from the segmented images Some feature reduction methods, which are Principle Component Analysis and Independent Component Analysis, provided that size of the feature vectors was decreased. All the modified vectors were given as inputs to various classifiers. Outputs of the classifier systems were evaluated and compared with each other. Accuracy values of proposed algorithm results were computed. Consequently, it is possible to use in real application of GPR devices.

Introduction

Ground Penetrating Radar (GPR) is a type of radar, provides information for many subsurface investigation including buried object detection, preliminary research for archaeological excavation, inverse scattering, and tomographic imaging. Besides, GPR is one of the non-destructive methods which is preferred for oil reserves and natural gas exploration, conduit and pipe localization. Sensors or antenna arrays can be used for underground imaging depending on the variety of applications. The GPR imaging technique is based on the transmitting and receiving of the electromagnetic wave in the field at a few meters depth with specific steps at the centimeter level. In fact, GPR uses wave reflections in regions where dielectric discontinuities exist in underground. Dielectric discontinuities can be formed by different soil layers, buried objects or clutter. The acquisition of information about the relevant region is per-formed depending on the time delay and phase calculation of the radar pulse. Neglecting wave refraction on the ground surface during inverse calculations of circular wave propagation can lead to incorrect configuration of the GPR images and incorrect detection of the target position. GPR B scan images consist of a combination of 1-D GPR A scan signals. GPR response is defined as combination of attenuation level in each depth index at measurement point. This response is called as 1-D GPR A scan signal. 2-D GPR B scan image is obtained by joining the received GPR A scan signal side by side in a specific step sequence. 3-D GPR C scan image is obtained by using 2-D GPR B scan images. In GPR B scan images, buried objects appear as hyperbola. The types of hyperbola differs in accordance with the material, shape and depth of buried objects.

The literature consists of few work related to detection and classification of buried objects patterns. Shihab et al. [1] used curve fitting method on the hyperboles in the GPR B scan image to determine the radius of buried cylindrical objects. Capineri et al. [2] applied Hough transform to segment linear lines owing to electrical discontinuity and determined the location of the buried objects. Caorsi et al. [3] estimated the properties of the buried cylinder material using artificial neural networks. Dolgiy et al. [4] used to determined radius of buried object via the weighted least squares method, the recursive Kalman filter technique and the maximum likelihood method. A scan signals, obtained from the reflected waves due to the dielectric difference, were separated into different frequency components to detect the existence of the object underground plane [5]. Delbo et al. [6] proposed a cascade system consisting of wavelet transform and fuzzy classification methods. Thanks to wavelet transform, the noise in the GPR B scan image is reduced and then, the hyperbolas in the image are segmented by fuzzy k means algorithm. Preprocessing algorithms were applied to reduce the noises around hyperbolas in GPR images by Gamba et al.  Afterwards, GPR images were interpreted by artificial neural networks and detector design was performed [7]. Al-Nuaimy et al. [8] performed preprocessing algorithms to reduce undesired noise on the image. Artificial neural network was successful for segmentation of GPR images. At the end, Hough transformation is applied to identify hyperbolas. Effects of different noise levels on GPR images were analyzed over detection of embedded objects by using artificial neural networks [9]. Rossini et al. [10] proposed an interpolation based wavelet transform method for the detection of underground objects. The high-level statistical properties obtained from the GPR B scan images are inputs of the artificial neural network and the earth surface and buried object are distinguished in the output layer [11]. Gamba et al. [12] have proposed two different methods, namely template matching and artificial neural network, for the detection of underground objects in small dimensions. Falorini et al. [13] eliminated found noise on GPR images by means of preprocessing algorithms and then, determined the same phase polarity point using 1-D gradient filter. In this way, the location of the target objects were detected. Dell’Acqua et al. [14] implemented 3-D Radon transformations on GPR B scan images to remove buried fake targets. Gurbuz et al. [15] performed the detection of the liner patterns on the GPR images by 2-D radon transformation. Borgioli et al. [16] estimated the location of the buried object by modifying the Hough transformation with weight factors. Pasolli et al.  [17] classified hyperbolic and linear patterns with support vector machines by formulating them with genetic optimization descriptor.

This study is divided into three categories in terms of stages. Initially, noise on the GPR B scan images are diminished with wavelet transform and the images are segmented by the Otsu thresholding method. Hough transform is used to separate the buried object and the ground plane. GLCM, RLM and ACF features are extracted from the segmented images.  Thanks to various classification algorithms, the shape of the buried objects is estimated. The remaining parts of this paper is organized as follows. The image processing and feature extraction methodologies are described in Section II. Feature reduction and classification methods are introduced in Section III. The experimental results are summarized in Section IV. Conclusions are existed in Section V.

Conclusion

In this study, a comparative analysis system of GPR images is presented. The proposed system allows to determine the buried of embedded objects. After the pre-processing and segmentation stages, different types of features were extracted and classification was performed. It has been tried to obtain the highest accuracy detection rate among proposed classifiers. Despite the reflections and overlaps in GPR images, the experimental results show up to 91.7% detection performance. The experimental results show that the proposed system has a high accuracy rate in terms of estimating the object shape even if buried objects are close to each other. This performance also depends on the number of GPR images in order to properly train the classifier in the proposed system.

1,297 total views, no views today

Download PDF File

About the author: admin