Elsevier

Information Sciences

Volume 196, 1 August 2012, Pages 73-96
Information Sciences

A retrieval system from inverse synthetic aperture radar images: Application to radar target recognition

https://doi.org/10.1016/j.ins.2012.01.049Get rights and content

Abstract

This paper presents an approach to aircraft target recognition using Inverse Synthetic Aperture Radar (ISAR) images. The goal of this work is to develop a robust algorithm to add Automated Target Recognition (ATR) capabilities to extract efficient feature vectors and ensure a comprehensive recognition process.

In the first part on this paper, we used a database of ISAR images reconstructed from anechoic chamber simulations in order to extract efficient feature vectors. In the second part of this work, we proposed a recognition architecture to perform recognition tasks and provide a human operator with useful information for target recognition tasks.

Several kinds of descriptor can be used to acquire information about target characteristics from radar signals and images. Indeed, a number of methods are currently used in automatic target recognition; notably based on ISAR imaging. However, target characteristic extraction from radar echoes remains a difficult task, and the methods are generally specific to either aircraft or ship recognition. In this work, we describe our approach to designing faster, more effective retrieval systems and a comprehensive architecture. Our approach uses global feature vectors for both ship and aircraft recognition.

Firstly, the global feature vectors are described by two types of descriptor which have shown to be efficient in radar target recognition. The first type defines target shape obtained by watershed transformation, facilitating interpretation for the human operator. The second type of vector descriptor is based on so-called polar signatures, obtained using the polar mapping procedure. The latter are highly discriminative and thus significantly improve and facilitate target recognition, leading to greater precision and accuracy. Secondly, we describe the retrieval architecture suitable for the second type of vector descriptor, which checks invariance in relation to target rotation and scale. It is, in itself, more efficient and processing time at the retrieval step is reduced and controlled by the human operator. Finally, in order to validate our proposed feature vectors and architecture, the Support Vector Machine (SVM) classifier will be implemented; the results of which will be tested and presented in the last section of this paper.

Introduction

Automatic Target Recognition (ATR) systems are an important part of modern military strategy. Correct recognition of military targets such as aircraft, naval ships, and missiles is indispensable to be able to attack the target accurately. Nowadays, radar and radar signatures are widely used for recognition purposes. However, several kinds of radar signature can be applied to acquire information about the target characteristics. For example, one can use modulation in the radar echoes produced by the rotating parts of a target such as propellers, rotors, or the compressor blades of a jet aircraft. Another option would be the extraction of geometrical parameters or characteristics of a target that can be obtained from a radar image. These characteristics can be obtained in one, two or n radar reflections in range. This single radar reflection is called a radar range profile. Under certain circumstances, the information about the target’s motion perpendicular to the line-of-sight can be extracted from a bi-dimensional image. That is why several techniques have been investigated in this context; several pieces of research have particularly focused on ISAR techniques and Automatic Target Recognition (ATR). These investigations typically use algorithms such as the Inverse Fast Fourier Transform (IFFT) and super-resolution radar imaging (for example MUSIC-2D and ESPRIT-2D algorithms). Several advances concerning ISAR image classification can be found in [16], [20], [23], [25], [29].

In the field of feature extraction, the main problem is to obtain feature vectors that have the invariance property relative to the scale and rotation of the ISAR image. For example in ship recognition using ISAR images, some feature vectors use a heuristic approach to estimate motion and geometric characteristics such as location of breaks in the superstructure profile or the number of major uprights [20], [2]. Such feature vectors are thus hardly applicable to aircraft recognition. In order to resolve this difficulty and to change the target scale, a number of theories have been proposed [2], [8]. A semi-automatic/automatic recognition scheme is proposed in this paper as an alternative in order to overcome operator limitations resulting from target recognition difficulties and to provide a useful tool to aid decision making.

Several feature vectors are desirable for interpretation and analysis in order to recognize different targets and therefore help a human operator to perform target recognition tasks correctly. The main strength of target recognition is that it uses target shape extracted from 2D-representations (ISAR images). We draw upon 2D ISAR images specifically because they furnish particularly detailed information about the geometry of the target (aircraft, naval ship, missile, etc.). The methodology used to design the complete processing chain from the acquisition step to the recognition (classification) step is based on the artificial intelligence approach. This process is known as Knowledge Discovery from Data (KDD) [14] which we have adapted to the radar target recognition system [29].

In this study, we put forward two main propositions. The first proposition is to provide the human operator with better information about target shape. Section 5 is devoted to the presentation of details and results concerning this proposition.

The second proposition presents other features that can be exploited to achieve satisfactory target recognition. In this proposal, processing time is reduced by using two approaches. The first approach is based on the processing of partial polar signatures in several steps as illustrated in Section 8. The second method is the use of non-supervised classification in order to optimize the training database and thus reduce matching time as presented in Section 9.

Section snippets

Automatic target recognition steps

Generally, four steps are involved in Automatic Target Recognition (ATR). They are data acquisition, data pre-processing, data representation and data classification for decision making. However, when the KDD methodology is applied to semi-automatic target recognition, another terminology is used for these steps: data acquisition, data preparation, Data Mining (DM), and the evaluation and decision making step. In the first step, data (signals) are collected from an anechoic chamber which is

Simulation data

A radar data acquisition system was studied using results obtained during tests conducted in ENSIETA’s anechoic chamber. This chamber facilitates the taking of real measurements and allows good control of the target-radar configuration. Thus, the human operator’s interpretation and control are made easier. The experimental setup is shown in Fig. 3.

To construct our simulation ISAR image database, we used eleven reduced-scale (1/48) aircraft models: F-104, F-117, Tornado, Harrier, A-10, F-14,

Shape extractor

The aircraft shape is one of the best features because it can give the human operator initial information on the unknown target. For this purpose, we chose to extract the shape from the ISAR image using the watershed method. Therefore, we used several processing steps.

In the first processing step, we segmented the ISAR image. In this segmentation process, the target response is extracted from the background which can include noise and clutter. Furthermore, in the anechoic chamber, we assumed

Shape descriptors

For shape representation, we used the contour-based methods that need extraction of boundary information. Therefore, for generic purposes, this kind of shape representation is necessary to accomplish the classification task. The goal of classification is to assign a new target to a class from a given set of classes based on the attribute values of these targets. The shape descriptors can be used in this way. In this section we describe Fourier descriptors and our classification scheme. Other

Results and discussion

In the simulation results, each target was represented by 162 ISAR images of 256 × 256 grayscale pixels. The initial database was divided into test and training databases. From each ISAR image, we computed a shape template that contained the compressed Fourier descriptors. In the results below, the size of each template was 20 elements (normalized boundaries = 80 points, Fourier descriptors df = 40 and PCA(df) = 20). It is important to take into account the size of training database. Some techniques

Polar descriptors

In this step, our aim was to construct the Polar Image (PI) using the Polar Mapping (PM) principle [16]. In the same principle, Log-Polar Mapping (LPM) can be used in this framework, but PM has a finer grid than LPM [16] because of its sampling interval in r-direction and the same sampling interval irrespective of the radius (see Fig. 2). The PM is shown in Fig. 21.

If we applied polar mapping to the initial image I(xi, yj) as defined in Section 5, we obtained an image called polar image Ip(rm, αn

Retrieval system for recognition

The classification scheme is based on a hierarchical architecture using three classifiers C1, C2 and C3. Each classifier is based on one part of the signature except C3 which uses V3 and V4 vectors as illustrated in Fig. 24.

After computing the normalized feature vectors Ir, Iα, and dfα from the polar image Ir,α, we constructed the training and test databases. In this step (Step2) it was possible to use some non-supervised classification methods. The authors in [16] uniformly sampled each class

Experiment results

To reduce the computation time, we used η < λ < 100%. Of course, the determination of λ and η values is a critical issue to guarantee a robust performance and to reduce time matching. However there is no systematic way of choosing them. Hence, we chose small values as desirable values. The idea was to reduce the size of the training database by using the Ward method in order to reduce the computation times and to improve classification accuracy.

In all the simulation results given below, we used the

Support vector machines

The Support Vector Machine (SVM) classification is based firstly on the training database to produce a model which is able to predict a label of the unknown images stored in the test database. This step is commonly called SVM learning. Given a labeled training vectors set {xi, yi}i=1,…,l with xi  Rn the template of Ir and dfα vectors and yi  {1, −1}l, support vector machines [9] require the solution of the following (primal) optimization problem:minw,b,ξ=12wTw+Ci=1lξis.t.yi(wTzi+b)1-ξiξi0,i=1,,l

Conclusion

The methods proposed in this paper have shown their adequacy for semi-automatic and/or automatic target recognition. In order to provide helpful information to a human operator, the target shape is extracted by watershed segmentation and the preliminary results are given. In this framework, some further processing and descriptors from shape information are really necessary to achieve feature vectors in order to obtain a satisfactory rate of correct recognition.

The retrieval system based on

References (36)

  • M. Bouet, A. Khenchaf, H. Briand, Shape representation for image retrieval, in: International Multimedia Conference,...
  • F.S. Cohen et al.

    Classification of rotated and scaled textured images using gaussian markov random field models

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1991)
  • C. Cortes et al.

    Support-vector network

    Mach. Learn.

    (1995)
  • S. Derrode, R. Mezhoud, Reconnaissance de formes par invariants complets et onvergents; application à l’indexation de...
  • H. Digabel, C. Lantuéjoul, Iterative algorithms, in: J.L. Chermant (Ed.), 2nd European Symposium on Qualitative...
  • R.O. Duda et al.

    Pattern Classification

    (2001)
  • B. Efron et al.

    An Introduction to the Bootstrap

    (1993)
  • W.J. Frawley, G. Piatetsky-Shaprio, C.J. Matheus, Knowledge discovery in databases – an overview, AAAI/MIT Press, 1991,...
  • Cited by (24)

    • Multivariate copula statistical model and weighted sparse classification for radar image target recognition

      2020, Computers and Electrical Engineering
      Citation Excerpt :

      To characterize the ISAR and SAR images, several methods has been proposed. Among them we distinguish the cropping, the down-sampling, the Histogram of Oriented Gradients (HOG) [2], the Fourier descriptors [1], the scale-invariant feature transform (SIFT) method [3], the local binary pattern (BLBP) [4] and so on. The study of radar images on the wavelet domain is a often used approach.

    • Combination of global and local filters for robust SAR target recognition under various extended operating conditions

      2019, Information Sciences
      Citation Excerpt :

      Interpretation of synthetic aperture radar (SAR) images has been widely studied for both military and civilian applications [19,21,33,36].

    • Similarity preserving multi-task learning for radar target recognition

      2018, Information Sciences
      Citation Excerpt :

      The work in radar automatic target recognition (RATR) area, according to the employment of the type of wideband radar signals, mainly focuses on high-resolution range profile (HRRP) recognition [5,6,10,14] and synthetic aperture radar (SAR) image or Inverse SAR (ISAR) image recognition [17,19,28,32,33].

    • Gravitation based classification

      2013, Information Sciences
    View all citing articles on Scopus
    View full text