Target detection of ISAR data by principal component transform on co-occurrence matrix

Abstract

Issue of automated target detection in ISAR can be stated as what features enhance objects of interest from the rest of the data. Much experimentation done in this area have used Fourier transforms for preprocessing the raw signal data. Generally the ISAR data are comes with a matrix of complex number values and therefore intuitive logic appears to favor a Fourier transform. A hypothesis was made that a Fourier transform in preprocessing may mask some data that could be part of feature used to threshold the object from background. Thus a trial was done on MATLAB simulated ISAR data to see if such data can be transformed into a matrix to visualize objects by preprocessing with principle component transform followed by some modification conventional thresholding techniques i.e. gray level co-occurrence matrix. Since it would be difficult to do so in complex valued matrices, these matrices had been decomposed to real valued and the imaginary valued matrices separately. Advantages of simulated data were that variables could be defined and changes in preprocessing transform and thresholding result could be compared with significant accuracy before a trial with actual performance of ISAR imagery. The preliminary result in this paper does show that preprocessing transform need not be Fourier. Principle component transform may bring about features that enhance thresholding values for Automatic target detection. Thresholding in conventional methods is done by finding a fixed value to create a binary image highlighting the object. In the modification proposed here single value thresholding objects and then spatially locating the object in a binary matrix may circumvented.

Introduction

Target detection is essential for target interpretation and analysis in Inverse Synthetic Aperture Radar (ISAR) imaging system. In ISAR, the target rotates and the radar is stationary and the target images can be obtained by transmitting wideband signals, and high cross range resolution is obtained by coherently accumulating number of echoes from different aspect angles. The goal of ISAR imaging system is to detect the targets particularly for surveillance. The conventional target detection system for ISAR images consists of the following stages: fast time filtering, slow time filtering, compression and decompression for focused FFT and IFFT response, Anti-aliasing and Matched Filtering. This paper proposed an algorithm for target detection for ISAR images which is based on co-occurrence matrix. Co-occurrence matrix is the statistical approach for texture representation and first introduced by Haralick et al. using the grey level co-occurrence matrix (GLCM) (Haralick et al., 1973). Co-occurrence matrix has been used to extract structural similarities between the objects (Mita et al., 2008, Jing Yi Tou et al., 2009), for classification (ZHU Le-Qing, 2010) and for segmentation (Corneloup, 1996). Co-occurrence matrix method also used by Clausi and Jerniganl (1998), where they proposed an improvement on the GLCM by presenting a grey level co-occurrence linked list (GLCLL) structure that stores the non-zero co-occurring probabilities in a sorted linked list. Rignot and Kwok (1990) have analyzed SAR images using texture features computed from gray level co-occurrence matrices. Threshold has been determined by using entropies, global, local and joint from co-occurrence matrix (Park et al., 2011, Mark et al., 1995, Chang et al., 1994).

A co-occurrence matrix captures the spatial dependence of contrast values, depending on different directions and distances specified. For a given matrix A with spatial dimension m × n with L gray levels G = 1, 2, 3, …, L. The gray levels of co ordinate (x, y) is denoted by $A(x\text{,}y)\in G$, the co-occurrence C of A is an L × L matrix

$$C=[f_{\mathit{ij}}{]}_{L\times L}$$

which contains the transition of gray levels with its adjacent gray levels. For the gray levels (i, j) the (i, j)th entry of the co-occurrence matrix C, $f_{\mathit{ij}}$ is defined as

$$f_{\mathit{ij}}={\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{i=1}^m}\sigma (m\text{,}n)$$

$\sigma (m\text{,}n)=1$ if $\{f(m\text{,}n)=i\text{,}f(m\text{,}n+1)=j\text{and}/\text{or}f(m\text{,}n)=i\text{,}f(m+1\text{,}n)=j$ $\sigma (m\text{,}n)=0$ otherwise

The probability of the occurrence is defined as

$$p(i\text{,}j)=\left[\frac{f_{\mathit{ij}}}{\sum f_{\mathit{ij}}}\right]$$

As such, the co-occurrence matrix can better expose the underlying nature of texture than can a Fourier description. This is because the co-occurrence measures spatial relationships between brightness, as opposed to frequency content. This clearly gives alternative results.

Many feature vectors has been computed from GLCM (Haralick et al., 1973, Tuceryan and Jain, 1998). Falconer et al. (2006) measured the power spectral density (PSD) of a variety of objects and used the differences in the shaping of the PSD (kurtosis and energy band) to differentiate between targets and infer the activity level (i.e., resting versus moving) of human targets. Sabatini and Colla (1998) used wavelet transforms to remove the high-frequency components and reconstruct the signal; the error between the original and reconstructed signals was then used to compute a threshold for discriminating between targets. López-Estrada and Cumplido (2009) used co occurrence matrix to evaluate the cluttered environment in a given image.

In detection module, firstly the raw data are processed for target detection by using co-occurrence matrix of the gray scale image. Principal component transform (PCT) using covariance has been done by taking co-occurrence matrix as input. Furthermore,it has been noticed that after applying the proposed method number of targets presents in the environment are clearly identified.

The paper is organized as follows; In Section 2 Methodology for capturing and preprocessing of ISAR data and algorithm for target detection has been given. Section 3 represents with all the results and discussion. Finally, conclusions are included in Section 4.