Comparative analysis of UWB deconvolution and feature-extraction algorithms for GPR landmine detection

In this work we developed target recognition algorithms for landmine detection with ultra-wideband ground penetrating radar (UWB GPR). Due to non-stationarity of UWB signals their processing requires advanced techniques, namely regularized deconvolution, time-frequency or time-scale analysis. We use deconvolution to remove GPR and soil characteristics from the received signals. An efficient algorithm of deconvolution, based on a regularized Wiener inverse filter with wavelet noise level estimation, has been developed. Criteria of efficiency were stability of the signal after deconvolution, difference between the received signal and the convolved back signal, and computational speed. The novelty of the algorithm is noise level estimation with wavelet decomposition, which defines the noise level separately for any signal, independently of its statistics. The algorithm was compared with an iterative time-domain deconvolution algorithm based on regularization. For target recognition we apply singular value decomposition (SVD) to a time-frequency signal distribution. Here we compare the Wigner transform and continuous wavelet transform (CWT) for discriminant feature selection. The developed algorithms have been checked on the data acquired with a stepped-frequency GPR.