Clutter reduction in ground penetrating radar images using morphological component analysis

Ground Penetrating Radar (GPR) is an important technology that is used for imaging the subsurface. Generally, GPR system consists of a transmitting antenna, a receiving antenna, a control unit and a displaying screen. GPR sends electromagnetic waves to subsurface with the transmitting antenna and reflected signals are collected from the receiving antenna. At the same time, the reflected signals are recorded and then processed with signal processing algorithms and finally processed data are displayed in the screen. GPR has been used in many applications: tunnel linings, geophysical investigations, urban infrastructure investigations, pipes and cable detection, archaeological investigations, road condition survey, military defence applications and so on. Thus, hardware and software properties of GPR varies according to application being used. In this work, GPR is used for military applications: the problem of detection of buried landmines. Detection and removal of landmines is a serious problem affecting civilians and soldiers. It is estimated that about 110 million landmines are buried in 64 countries around the world and that 26,000 people a year are killed or maimed by a landmine. There are two types of landmines: anti-personnel and anti-tank. Anti-personnel landmines are designed for use against humans, as opposed to anti-tank mines which are designed for use against vehicles. Anti-personnel landmines are equipped with casing that protects mine from environment. Early landmines had casing made of steel and aluminium but these mines are easily detected by metal detectors. After the use of metal detectors for landmine detection, mines manufacturers produced new landmines which include minimum metal content. The performance of metal detectors were degraded after the production of these landmines and GPR is started to be used for the detection of landmines with little or no metal content. Although GPR offers the promise of detecting landmines with little or no metal content but landmine detection via GPR has been a difficult problem. It is well known that target detection process in GPR is highly affected by the clutter. Clutter can be caused by breakthrough between the transmitting and receiving antennas, reflection from the ground which is called ground bounce and scattering response from non-mine objects (roots, small rocks, non-uniform terrain and so on). Since the targets are buried near the surface which are consists of minimum metal contents, clutter suppresses target signal because amplitude of reflected signal from the target is weaker than ground bounce. Briefly, clutter prevents to detect landmines and cause false alarms in the non-target region. In addition, the performance of GPR system is affected from soil conditions, temperature, weather conditions and varying terrain. Clutter reduction techniques has gained importance to achieve high detection rates and to decrease false alarm probabilities. In GPR system, the reflected signal is composed of clutter, target and system noise. As the system noise has less importance compared to the other components, clutter reduction algorithms aim to decompose the reflected signal as target and clutter. There are three different approaches to remove clutter in the literature. These techniques are based on modeling the clutter or the target and subspace based statistical methods. In the first approach, clutter or background model is estimated from non-target region and this estimated clutter is subtracted from all GPR image. At the end of this process clutter is removed from GPR image and the image includes only target information. Background (Mean) Subtraction (BS), Median Subtraction (MS), Moving Background Subtraction (MBS), Moving Median Subtraction (MMS), Kalman, Wavelet and Likelihood Ratio Test (LRT) are methods which models clutter. This approach needs information about target-free region and the performance of these methods degrade with varying terrain conditions. In the second approach, target is modeled and then searched in the GPR image. This approach consists of the training and the test stages. Training stage requires too many target data which are taken from different soil conditions, different terrain, large variety of landmine types. After the collection of target data, feature extraction process is applied to these target data. Extracted features are trained with learning algorithms and targets are modelled. In the test stage, feature extraction process is applied to suspicious region in the GPR data and extracted features are compared to target model. After the comparison, the algorithm decides whether the observed region contains target or not, so clutter is separated from target. Especially, this approach is appropriate for anti-tank landmines because GPR signature of anti-tank landmines are suitable for modeling. Polynomial fitting, Histograms of Oriented Gradients (HOG), Hidden Markov Models (HMM), Edge Histogram Descriptors (EHD) are algorithm which are use this approach. In addition, these methods' computational burden is very high so these methods are applied only to suspicious regions to separate target from clutter and also need auxilary algorithms to specify suspicious regions. Subspace based statistical methods are the third approach that are widely used methods in literature for clutter reduction in GPR data. In this approach, GPR data is decomposed into sub images corresponding to clutter, target and background components. Singular Value Decomposition (SVD), Principal Component Analysis (PCA) and Independent Component Analysis (ICA) are used for clutter reduction in this approach. Since in GPR the clutter is much stronger than the target component, it can be reconstructed by the use of the eigenvector corresponding to the largest eigenvalue of the correlation matrix of the GPR image in SVD method or the first principal component in the PCA based method. Similarly, in ICA based method the components are reconstructed using corresponding independent components. Bell-Sejnowski, Molgedey-Schuster, FASTICA, Joint Approximate Diagonalization of Eigenmatrices (JADE) and Kernel-ICA (KICA) algorithms have been developed to implement the ICA method. In this thesis, JADE and FASTICA algorithms are used to remove clutter from GPR images. There are a lot of publication in literature that compare these subspace methods with each other and other approaches. In general, ICA based methods outperform other methods for clutter reduction in GPR images. It is known that these methods decompose GPR image with selecting components and clutter is reconstructed using the first largest component but it is not known exactly which components includes target information. There are some approaches to calculate which components contain target information but achieved results are not satisfactory and component selection process for the determination of the target component is still an open area of research. In this thesis work, the first time Morphological Component Analysis (MCA) method is proposed for clutter reduction in GPR images. MCA is an image decomposition method based on sparse representations and morphological diversity and it is proposed to decompose images into texture and piecewise smooth (cartoon) parts. The basic idea presented in MCA algorithm is the use of two appropriate dictionaries, one for the representation of textures and the other for the cartoon parts. In sparse signal processing area, dictionary is defined as a set of signals and constructed with one dimensional signals or two dimensional image parts which are extracted from images and called "patch". Each element of dictionary is referred as "atom". There are two types of dictionaries: analytic dictionaries and learning based dictionaries. In analytic based approach, dictionaries are constituted with using analytic functions like cosine, sine. In the learning based dictionaries, patches are extracted from a set of images and these patches are processed via learning algorithm and learning based dictionary is constructed with using learned (processed) patches. MCA assumes that an image is a linear mixture of several morphological components where each of them can be sparsely represented with an appropriate dictionary which is inefficient to represent sparsely the other component. This is a key assumption for the success of the MCA algorithm, so the success of MCA algorithm depends on the selection of the dictionaries. Therefore, choosing a suitable dictionary is an important step with two criteria: i) dictionaries must be mutually incoherent and, ii) dictionaries must represent each related component sparsely. Once the dictionaries are identified, MCA finds the components by successive iterative thresholding. In each iteration of MCA algorithm, each component is represented by the related dictionary and sparse coefficients. Sparse coefficients are calculated to represent each component with related dictionary and then these calculated coefficients are thresholded. MCA uses different thresholding methods and thresholding strategies. Thresholding methods are applied to coefficients while thresholding strategies calculate related threshold in each iteration. GPR images include linear mixture of clutter and target components and separation of these components from each other is challenging. In this study, GPR images are decomposed into clutter and target components by the use of MCA. To successfully decompose GPR image into clutter and target components, suitable dictionaries for each type of image-content must be selected. Before the dictionary selection process, GPR images are investigated and it is observed that clutter components include lines or line-shaped parts and targets produce isotropic structures. It is reported that isotropic structures can be efficiently represented by wavelet dictionaries while for global lines curvelet or ridgelet dictionaries are more appropriate candidates. Thus, the clutter and target components are sparsely represented with curvelet and Undecimated Discrete Wavelet Transform (UDWT) dictionaries, respectively. A new clutter removal method based on MCA is applied to four two dimensional GPR images (B-scans). These four B-scans are obtained by using four different scenarios and each scenario is composed of metal targets and as well as plastic targets with low metal content and is performed with different soil conditions and soil types. Obtained results show that MCA is efficient and appropriate method to remove clutter from GPR images. On the other hand, SVD, PCA and ICA methods are applied to these B-scans for comparison purposes. Obtained results show that MCA outperforms these state-of-the-art clutter removal methods.