Study of Feature Extraction and Classification Algorithm for HRRP
|School||Xi'an University of Electronic Science and Technology|
|Course||Signal and Information Processing|
|Keywords||Radar high resolution range profile zero phase represent bispectra multiple polarization information incremental algorithm power method principal component analysis linear discriminant analysis subspace projection algorithm kernel principal component analysis|
Radar high resolution range profile (HRRP) represents the projection of the complex returned echoes from the target scattering centers onto the radar line-of-sight (LOS). It contains the target structure signatures, and it is easier to get compared with SAR and ISAR images. Thereby radar HRRP target recognition is a promising technique. Efficient feature can obtain high recognition performance with low computation burden; Fusion of multiple kind of different features can improve the recognition performance too; and in realistic application, we can’t obtain all the training samples in once time, so it’s necessary to develop efficient incremental algorithm. The contribution of this dissertation is concentrated on extraction of shift-invariant feature, fusion of multiple polarization and multiple feature; and derivation of incremental algorithm. The main contents are summarized as follows:1) For the shift variant and dimension reduction in HRRP recognition, the zero phase represent is first introduced to make the HRRP self-aligned, then the aligned HRRP is compressed by discrete cosine(sine) transformation, the new feature has low dimension and good expansion ability; Because the single range cell often contains many scatters, so its amplitude is sensitive to target aspect and hard to be used, but when we sort the HRRP according the scatter’s amplitude, a new shift-invariant feature is obtained, the experiments testified amplitude contains useful discrimination information too.2) Bispectra is a widely used shift-invariant feature, but the high dimension restricts its realistic application. Two methods are presented for its dimension reduction. The first one is low frequency bispectra, through experiments we find most points with high discrimination power located at low frequency region, so we use the low frequency part of bispectra as the recognition feature; the second one is based on the singular value decomposition of bispectra matrix, the singular values and singular vectors corresponding to large singular values are used as the reduced features.3) Multiple polarization HRRP can provide more target information, and different feature of HRRP contains different discrimination information, we fuse all these information with D-S theory and promote the recognition performance much; Nearest feature line (NFL) classifier can relax the conflict between the recognition performance and sample number, but it has a high computation complexity. So a local aspect NFL is promoted with lower complexity, with the data lenghthening technique, the recognition performance is promoted again.4) With the detailed analysis we point out that the CCIPCA algorithm is a kind of online power method, and a new incremental BDPCA algorithm is presented based on it; Two simplified subspace projection algorithms are presented with high performance, the first one use the character of eigenvector and the second one bases on approximated covariance matrix which is composed with large eigenvalues and corresponding eigenvectors; and the incremental LDA algorithm based on power method has low complexity and high estimation accuracy.5) Kernel PCA can use the kernel function project the feature from low dimension space to high dimension space, so the target is easier distinguished in that space, but its complexity is dramatically increased with the sample number. We first present a new KPCA algorithm with more concise form and then an incremental KPCA algorithm is presented based on it too.