Multi-view features-based HRRP classification via sparsity preserving projection

摘要

Feature extraction is an important step in radar high-resolution profile (HRRP) recognition area. Traditional feature extraction methods are constrained to the original signal itself without taking advantage of abundant information exists in multiple views. In this paper the multi-view sparsity preserving projection (MVSPP) method is proposed for feature extraction of HRRP. The core of this method is taking sparsity preserving projection to reduce dimensions of multi-view features and introducing kernel trick to ensure their similarities. Based on the characteristic of SVM classifier we propose a new combined kernel function instead of the traditional linear kernel function. To obtain the optimal solution, MVSPP adopts an iterative procedure. Experiments based on measured HRRP data verify the advantage of the proposed method. In addition, factors affecting the convergence speed of the algorithm are analyzed. Moreover, the performances of two kinds of kernel functions are compared.