Object Recognition from gpr images

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

1 of 7 

Object Recognition from GPR images

 

 

Executive summary of project 

The goal of this project consists in designing a system of automatic recognition of objects in images 

obtained with an impulse GPR. Development of such a system suggests the solution of the following 

problems: 

1.

 

development of methods for automatic object selection in images; 

2.

 

development of methods for object description by a system of features; 

3.

 

estimation of possibility of recognising given classes of objects from the obtained feature vector;  

4.

 

development of efficient recognition algorithms.  

The choice of methods for solving the first and second problems essentially depends on signal pre-

processing used in the GPR and methods for constructing GPR images. 

We are aimed to develop two groups of methods for object selection in an image. The methods of 

the first group take into account specific features of object patterns on the GPR images constructed without 

aperture synthesis and are based on wavelet-analysis. The methods of the second group are independent 

of specific features of GPR images and include image filtration and segmentation. 

Methods of object description by a feature vector also depend on the procedure of constructing 

radar images. In the context of the first group of methods for object selection, the object description is 

based on the wavelet-expansion of the radar image. For the second group, an object is described by the 

feature set able to characterize the object shape and texture. 

For object recognition, we suggest the use of tree-like statistical classificators. The advantage of 

such an approach consists in the fact that such classificators allow for estimating classification quality if the 

alphabet of classes to be recognized in known.  

 

Novelties

 

1.

 

The use of the two-dimensional wavelet specially designed for selecting and describing an object in the 

radar image. 

2.

 

The use of original segmentation algorithms for selecting objects in the radar image. 

3.

 

Description of object shape with the specially designed technique that includes expansion of the object 

contour in a set of orthogonal functions (in particular, these functions can be wavelets) and construction 

of a metric that offers a possibility of comparing arbitrarily shaped objects independently of object 

translations, rotations, and reflections. 

4.

 

Derivation of new tree-like statistical classificators (classification algorithms). 

 

 

Timescale

 

 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

2 of 7 

% of activity of scientists 

Names 

Year 1 

Year 2 

Vladimir Sazonov 

15 

15 

Aleksandr Vinogradov 

20 

20 

Natalya Verdenskaya 

20 

20 

Irina Ivanova 

20 

20 

Vladimir Chepelev 

35 

35 

Dmitrii Chikin 

40 

40 

Andrei Tereshin 

40 

40 

 

Description of the project 

(a)

 

Technical/scientific description 

The general scheme of designing the recognition system is the classical scheme shown in Fig. 1: 

 

Signal pre-processing and clutter suppression 

 

Data focusing 

 

Segmentation of the not-focused and focused images 

 

Feature extraction 

 

Classification 

Fig. 1. 

Either initial or pre-processed data can be used when solving the problems of object detection and 

object selection. Both approaches have their advantages and limitations. Indeed, dealing with the initial 

data, we loose no informative data; however, the object pattern and feature set describing the object will 

essentially depend on the procedure of getting the data. 

If  the data are the pre-processed data (for example, the data is the SAR image constructed from 

received signals), the information can be partially loosed. However, one can face with the opposite 

situation in which the image will have greater resolution and object description will be universal. 

Both approaches of working with data are suggested for the consideration. 

Signal pre -processing and clutter suppression 

In line with the commonly used methods of suppressing clutters from the ground surface (such as 

subtraction of the signal averaged over certain window), we plan to analyse the usability of other filters. In 

addition we will attempt to construct the model of the desired signal and the model of environmental 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

3 of 7 

clutters. With these models, we hope to be able of suggesting the procedure for selecting the desired signal 

against the background of environmental clutters. 

Data focusing 

We suggest to use data preprocessing according to the following scheme: 

-

 

synthesis of the complex radar hologram using quadrature signals. If radar is unable to furnish two 

quadrature signals, we hope to form the second quadrature signal by the Hilbert transform of 

received signal, 

-

 

inverse filtration, 

-

 

aperture synthesis. 

Non-focused image segmentation 

Depending on size, shape, and reflectivity (refractivity), an object form in the GPR image a pattern of 

several perfect or distorted hyperbolas or ellipses in time or space slices. We suggest the use of certain 

special wavelet expansion for selecting such patterns. With this goal in view, we will attempt to construct a 

two-dimensional wavelet function with backbone/valley of hyperbolic and elliptic forms. The scale of such 

a wavelet is governed by two curve parameters. The rate of backbone decay will form an additional 

parameter of the wavelet. The expansion of a GPR image in wavelet series with this base wavelet will offer 

a possibility of selecting a family of hyperbolas related to the same object. The set of parameters 

corresponding to this family, in turn, will offer object description by the set of features. 

Focused image segmentation 

This processing uses the standard filtration-segmentation-selection scheme. The first step—

filtration—is assumed to be performed according to standard procedures. For the second step—

segmentation (it is meant as selecting image segments homogeneous in the sense of texture)—, we suggest 

to use statistical algorithms of segmentation, which are modified algorithms for mixture separation. The third 

step—object selection—assumes selecting connection regions in the image, and these regions form the 

object pattern. Both of these procedures (filtration and segmentation) are based on pixel (image dots) 

probability distribution features. For this reason, these procedures allow multi-dimensional extensions 

simply by changing the dimension of the likelihood function. Here we mean 2D and 3D real-valued images. 

Processing of the complex-valued image require certain transformation of the above algorithms, because 

the likelihood function changes the number of parameters in this case. 

 

Object feature extraction 

For object description, we suggest to form the feature vector from an analysis of object shape, 

shapes of object constituents, and texture parameters of these elements. 

Despite vast literature about the describing and comparing shapes of arbitrary closed curves, this 

topic is far from being positively solved. However, this topic is urgent in recognising objects in an image. 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

4 of 7 

The solution that we suggest to consider in this project is as follows. An arbitrary closed contour is 

represented as a periodic complex function of real argument. This function is expanded in the series in 

certain system of orthogonal functions (in particular, in wavelets). Then, the function is approximated to a 

given error by a finite sum of expansion terms, so that the finite set of coefficients used in the finite sum 

describe the function to a given error. Such curve representations are compared using either the specially 

constructed metric invariant relative motions in the plane, or the standard Euclidean metric. In the case of 

the Euclidean metric, a specially transformed set of coefficients is used as curve description instead of the 

initial one.  

The shape of a three-dimensional object can be described using algorithms of object formation from 

sections or from expansions in a series in the multidimensional systems of functions. 

In addition, we plan to consider the possibility of constructing the complex object description that 

will include the images obtained for both polarisations.  

Standard procedures of decreasing dimension, such as the Karhunen-Loeve expansion, can be used 

to decrease the dimension of the feature vector. 

Classification 

At the first stage, we suggest to consider the problem of recognition with training. In this case, an 

additional step appears in the scheme—construction of training samples from experimental data. These 

samples are used for the formation of the alphabet of classes to be recognised in the developed system of 

features with and without using the corresponding procedures of decreasing dimension. The same training 

samples are used for estimating the possibility of recognising a given set of object classes. 

We suggest to design classification algorithms using statistical tree-like classificators. 

For estimating the possibility of classifying the given set of object classes, we suggest to use the 

procedures of automatic classification (such as Forel) that allow estimating the quality of resulting division. 

The final choice of the algorithms will be done after a half year of working with the IRCTR data and 

will be agreed with IRCTR. 

The possibility of real-time processing is mainly limited by the speed of algorithms. The speed of 

every algorithm can be estimated in terms of the number of required operations. In this case, the main 

portion of calculations is evaluated as the body of data to be processed multiplied by of the number of 

operations required to process a single dot of the image. The quality of fast algorithms using a small number 

of operations per one dot is lower than the quality of more complicated procedures, which are, naturally, 

slower. The choice of specific algorithm will be done at the stage of testing with actual data obtained from 

GPR. 

Thus, the speed of the processing procedure and, consequently, the possibility of its real-time 

implementation depend on the following basic factors: 

-

 

speed of the algorithm (the number of operations required to process a single dot of the image); 

-

 

body of data to be processed; 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

5 of 7 

-

 

required quality; 

-

 

the speed of the computer used. 

It is assumed that processing algorithms will use the digital GPR signal as the initial data. All the 

algorithms will have a set of parameters to be tuned. The parameters will be tuned during testing the 

algorithms with data obtained from the IRCTR video  impulse radar with consideration for basic GPR 

performances, such as radar bandwidth, sampling time, and acquisition parameters. 

The quality of algorithms will be examined using model received signals formed from the model of 

object reflectivity and the model of surrounding medium. With these models, we assume to form test 

images that will be used for estimating the quality of processing algorithms. Detection and false alarm 

probabilities will be estimated using actual samplings obtained from the IRCTR radar. The algorithms to be 

developed must have characteristics better that the characteristics of the energy projection detector. 

All algorithms will be implemented as the dynamically linked library (DLL). We assume to develop 

the algorithms using Delphi 7.0  as the main development medium.  We deliver the description of the 

algorithms, text of routines, description of the library and the examples of MATLAB calls of the main 

functions from the library. 

 

(b) Utilization 

As the main utilisation of the results of this project, we see the automation of searching for objects of 

relatively small size buried for relatively small depths, such as 

•

 

unexploded projectiles, landmines, and other ammunition; 

•

 

foreign objects of other origin and, in particular, organic inclusions; 

•

 

air-filled cavities of small size.  

 

The main purpose of this project consists in automation of deciding the availability of an object, 

which will offer a possibility of making the process faster and exploring large areas, possibly, without direct 

participation of personnel. 

(c) Applicants 

 

No 

Problems 

Applicants 

1 

Data pre-processing (clutter suppressing, designing of 

models for clutter and desired signal) 

Aleksandr Vinogradov, 

Andrei Tereshin 

2 

Image construction by the synthetic aperture (SA) 

method 

Vladimir Sazonov, 

Dmitrii Chikin 

3 

2D and 3D segmentation 

Natalya Verdenskaya 

4 

Feature extraction 

Irina Ivanova 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

6 of 7 

5 

Object recognition 

Natalya Verdenskaya 

Irina Ivanova 

Vladimir Chepelev 

 

Expertise of the applicants in the area of the project 

 

Aleksandr Vinogradov 

Graduated from Moscow Institute of Physics and Technology (MIPT), Department of General and 

Applied Physics (1970). 

Degree of Candidate of Sciences in Physics and Mathematics (1974). 

Thesis: "Effect of Backscattering Enchancement". 

Scientific title: Senior Scientific worker in Radiophysics. 

Knowledge: radiophysics, mathematics, wave propagation in random media, radar engineering. 

Experience: with Mints RTI from 1970, participated in many projects on designing radar systems as 

theoretician, one of the main designers of automatic system of blood cell recognition and counting. 

With MIPT from 1990 (assistant professor). Lectures: "Theory of diffraction" (1990-2000), "Wave 

propagation" (2000-present). Supervision. 

Publications: more than a hundred papers and scientific reports. 

 

Natalya Verdenskaya 

Graduated from Moscow State University, Faculty of Mechanics and Mathematics (1985). 

Degree of Candidate of Sciences in Physics and Mathematics (2002). 

Thesis: "Image segmentation algorithms and their application to designing automatic recognition systems in 

medicine and radar engineering". 

Knowledge: mathematical statistics, image analysis and image processing, pattern recognition. 

Experience: with Mints RTI from 1987, participated in some projects on designing radar systems as 

theoretician, one of the main designers of automatic system of blood cell recognition and counting. 

With MIPT from 2003 (assistant professor). Lectures: "Theory of image processing". Supervision. 

Publications: more than a fifty papers and scientific reports. 

 

Irina Ivanova 

Graduated from Moscow State University, Faculty of Computational Mathematics and Cybernetics 

(1987). 

Knowledge: mathematical statistics, image analysis and image processing, pattern recognition. 

Experience: with Mints RTI from 1987, participated in some projects on designing radar systems as 

theoretician, one of the main designers of automatic system of blood cell recognition and counting. 

Object Recognition from GPR images 

Project TUDelft/IRCTR en MIPT, printed: 20/03/2006, page: 

7 of 7 

With MIPT from 2003 (assistant professor). Lectures: "Theory of image processing". Supervision. 

Publications: more than a forty papers and scientific reports. 

 

Vladimir Chepelev 

Graduated from Moscow Institute of Physics and Technology (MIPT), Department of General and 

Applied Physics (2000). 

Knowledge: radiophysics, radar engineering, pattern recognition. 

Experience: postgraduate student MIPT. 

Publications: 7 papers. 

 

Dmitrii Chikin 

Student, MIPT, IV year. 

 

Andrei Tereshin 

Student, MIPT, V year. 

Publication: 2 papers.