Deep unfolding for clutter removal in ground penetrating radar /

Abstract

Ground Penetrating Radar (GPR) is commonly used for identifying subterranean items, particularly plastic landmines that have a small amount of metal, cavities and pipelines. Nevertheless, reflections emanating from subterranean targets are considerably impacted by the mess created by the direct connection between the antennas used for transmitting and receiving signals, reflections originating from the surface of the soil, and the scattering effect caused by objects other than mines, such as roots, gravel, or uneven terrain. The clutter is preventing to detect targets and decreases the performance of detection algorithms since it is dominating targets' reflections. For this reason, the clutter removal of GPR images plays a critical role to detect clearly underground objects. For decades, lots of methods are proposed for clutter removal from GPR images. The subspace decomposition techniques such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Singular Value Decomposition (SVD) can be used to decompose GPR images into distinct components corresponding to the clutter and the target based on their strength differences. Non-negative matrix factorization (NMF) can be applied to GPR images to extract meaningful features and reduce the dimensionality of the data. Robust PCA (RPCA) decomposes a matrix to low-rank and sparse components. Due to high computational cost SVD operation in each iteration to solve non-convex minimizing operation that increases run-time. Several efforts have been made to accelerate the process of SVD operations in field studies suchs as Go-Decomposition (GoDec) which uses bilateral random projections (BRP) to directly obtain the low-rank component. Also robust NMF (RNMF) has been proposed, which includes a sparse component in the NMF decomposition. Deep learning methods are used in GPR imaging detection target and clutter removal tasks. The Robust AutoEncoder (RAE) is unsupervised deep learning method that can serve as a viable alternative to RPCA when it comes to decomposing a matrix into low-rank and sparse components. RAE employs the l_1-norm to break down a sparse matrix, much like RPCA does. The following traditional methods have been proposed to reduce the high computational cost and working time: Learned Robust Principal Component Analysis (LRPCA), Convolutional Robust Principal Component Analysis (CORONA), and Go-Decomposition Network (GODEC). These proposed methods are referred to as Deep Unfolding Networks (DUN) algorithms. In general, DUN algorithms convert each iteration of iterative methods into a deep learning layer in order to reduce the aforementioned computational cost and working time. LRPCA expands the concept of deep unfolding from a limited number of iterations to an infinite number of iterations by utilizing a unique feedforward-recurrent-mixed neural network model. This approach allows LRPCA to improve the runtime performance during the testing phase compared to traditional methods, while also enabling the learnability of hyperparameters. On the other hand, CORONA is recognized for its fast algorithm, but its computational intensity is amplified due to the utilization of SVD in each iteration. However, a newly proposed method called GODEC-Net addresses this issue by employing BRP instead of SVD at each iteration. As a result, this modification significantly enhances the speed of regular operations for CORONA. In this thesis study, a dataset is created using GPR images obtained from real-world conditions. The dataset is divided into two parts: training and testing. In order to train the DUN algorithms, the training dataset included not only raw data from GPR images but also separate images for target and clutter. To achieve this, the training dataset is initially decomposed into its components using RPCA. The proposed algorithms are compared visually and numerically with SVD, RPCA, RNMF, GoDec, and RAE. According to the obtained results, our proposed algorithms yielded similar results to other methods and, in some scenarios, are observed to perform better in clutter removal. LRPCA, although faster in terms of processing speed compared to CORONA and GODEC-Net, it has been demonstrated that CORONA and GODEC-Net perform better in clutter removal. The aim of the thesis is to remove clutter in GPR images using deep unfolding networks, instead of computationally expensive and time-consuming algorithms. By doing so, it becomes more flexible to learn the iterative algorithm parameters that vary from one application to another using the proposed algorithms. This enables the realization of clutter removal in real-world applications.