New clutter removal methods for through obstacle target detection

Nowadays, ultra–wide band (UWB) radars are extensively used for target imaging, detection and localization, health care applications, indoor motion detection, civil engineering applications, etc. UWB radars have also found many applications for buried object detection problem or through the wall target localization problem due to their ability to penetrate most of the non-metal materials such as bricks, wood, various types of soil, concrete and reinforced concrete, etc. In this thesis, we focus on the buried object detection problem where the radar used for this specific application is called as ground penetrating radar (GPR). It is known that one of the most important problems in images collected from GPR (it is called as GPR image or B-scan) is clutter since it degrades the detection performance. The clutter in the GPR image can be arisen by a various causes such as direct-wave arrival, ground-bounce, other candidate target-like objects, environmental variations (soil type, material type, etc.), weather conditions (sun, rain, snow, etc.). Therefore, it is impossible to eliminate the causes of clutter thus the obtained GPR image contain inherently heavy clutter. This heavy clutter conceals and attenuates the signal strength of the buried objects thus removing the clutter effect in the GPR image will certainly increase the probability of detection performance of the methods. To compare the available clutter removal methods with the proposed ones, we need a sufficient size simulated dataset and real GPR dataset in order to validate the effectiveness of methods with real GPR images. The simulated dataset is required since it can be used to report quantitative results besides the visual ones. The simulated dataset in this thesis is constructed by gprMax free–software which it uses Yee's algorithm to solve Maxwells equations in 3D using finite-difference time-domain (FDTD) method. The simulated dataset contains 112 GPR images and these GPR images are constructed by using seven different soil types, 2 different objects, 2 different burial depths and 4 challenging scenarios. In addition, the popular real GPR data from Vrije Universiteit Brussel is used to validate the performance of algorithms in the sense of visual perception. There are two main groups of approaches for clutter removal methods for buried object detection in GPR images. The first group is subspace–based methods such as principal component analysis (PCA), independent component analysis (ICA), singular value decomposition (SVD) and morphological component analysis (MCA). PCA, ICA and SVD perform matrix decomposition on the GPR image using different matrix decomposition constraints. After the GPR image decomposes into multiple sub–images by these methods, the first sub–image (most dominant one) is accepted as clutter and the remaining ones construct the target components. MCA is listed in this group but it uses the trained dictionaries in the pre-processing step to separate the GPR image into morphological components such as clutter and target. The second group approaches are based on the multi–scale and multi–direction (multi–resolution) based methods such as curvelet transform (CT), nonsubsampled contourlet transform (NSCT), shearlet transform (ST) and etc. These methods are using different scaling and directional subbands to express the GPR image as a decomposition of coefficient matrices with multiple scales and directions. Then, the information found in different directions is used to separate the target from clutter. These methods are exploiting the intrinsic geometrical property of GPR image where the target has hyperbolic direction and the clutter has horizontal direction in GPR image. The most popular multi–resolution based method for clutter removal problem in GPR is CT since the optimized version of CT is publicly available with the name of fast discrete curvelet transform (FDCT). In this thesis, we propose methods for both group. For the first group, the non-negative matrix factorization (NMF) method is proposed which is extensively used in signal processing community for dimensionality reduction, feature extraction, source separation and classification purposes. This method is similar to other subspace–based methods however its constraint is different where all the elements in the decomposed matrices are non–negative and it is a reasonable constraint for clutter removal problem in GPR since the pixel intensity values are non-negative in GPR image. The clutter removal methodology of NMF is similar to PCA, ICA and SVD however the performance is better compared to them and the running–time of the proposed method is fast enough for real–time implementations. For the second group, multi–scale directional bilateral filter (MDBF) is proposed. This method uses bilateral filter (BF) method for multi–scale decomposition part and directional filter bank (DFB) for the multi–directional decomposition part. After the GPR image is decomposed into multiple directional subbands in different scales, the target and clutter components are found by using their inherent directional property and the horizontal information is removed in the decomposed subspace where it corresponds to the clutter. Then, the inverse transform is applied to the remaining subbands and the clutter-free GPR image is reconstructed. MDBF algorithm is proposed firstly and it is improved later by replacing multi–scale part (BF part) with neighborhood filter (NF). NF filter is similar to BF filter however it uses only one Gaussian kernel which depends on the range information and it is more reasonable for this problem. The additional spatial Gaussian kernel in BF does not contribute to the performance. Besides, eliminating one Gaussian kernel also reduces the computational complexity by half and help to increase the running–time of the MDNF method. Different metrics are proposed for the performance comparison of clutter removal methods in GPR images such as peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), receiver operating characteristic (ROC) and etc. Most of the performance metrics need reference (ground–truth) image however our simulated dataset does not have one. Therefore, ROC curves are used as evaluation metric which depends on the probability of detection (PD) vs. probability of false alarm rate (FAR) for quantitative analysis as well as visual results are presented. Both quantitative and visual results have shown that NMF has superior performance compared the other available popular subspace-based methods in literature and MDNF also outperforms the available and recently proposed multi–resolution based methods. In the overall comparison, MDNF show the best performance among all methods however it is not optimized as in FDCT thus the proposed NMF method is more suitable for fast implementations.