Performance evaluation of NMF methods with different divergence metrics for landmine detection in GPR

Ground-penetrating radar (GPR) is a non-destructive geophysical tool used to detect buried objects. However, detection of shallowly buried objects such as landmines is a challenging problem due to the inherent presence of the clutter. Various methods based on subspace decomposition or multiresolution analysis are proposed for clutter removal in GPR images. The recently proposed subspace based non-negative matrix factorization (NMF) method is similar to the other well-known image decomposition methods however it has different constraints such as all the elements in the decomposed matrices have to be non-negative which more appropriate for our problem. The method is based on the low rank approximation of the GPR image. Several divergence metrics/cost functions have been proposed in literature for NMF as a convergence criteria such as Euclidean (EUC) distance, Kullback-Leibler (KL) divergence and Itakura-Saito (IS) divergence. These metrics affect the performance of NMF during the clutter removal process. To find the most suitable divergence metric in NMF for GPR clutter removal problem, a simulated dataset is constructed by using gprMax free software. The GPR images in our constructed simulated dataset have also the ground-Truth images and represent challenging scenarios. Therefore the quantitative results are given in addition to visual results which is hard to obtain in the real GPR measurements. For the quantitative analysis, the peak signal-To-noise ratio (PSNR) and the structural similarity (SSIM) performance metrics are used since the ground-Truth images are available. Both the quantitative and visual results show that NMF with KL divergence outperform the other divergence metrics for GPR imaging.