Hierarchical Recognition System for Target Recognition from Sparse Representations

Abstract EN

A hierarchicalrecognition system (HRS) based on constrained Deep Belief Network(DBN) is proposed for SAR Automatic Target Recognition (SAR ATR).

Asa classical Deep Learning method, DBN has shown greatperformance on data reconstruction, big data mining, andclassification.

However, few works have been carried out to solvesmall data problems (like SAR ATR) by Deep Learning method.

In HRS,the deep structure and pattern classifier are combined to solve smalldata classification problems.

After building the DBN with multipleRestricted Boltzmann Machines (RBMs), hierarchical features can beobtained, and then they are fed to classifier directly.

To obtain morenatural sparse feature representation, the Constrained RBM (CRBM) isproposed with solving a generalized optimization problem.

Three RBMvariants, L1-RNM, L2-RBM, and L1/2-RBM, are presented andintroduced to HRS in this paper.

The experiments on MSTAR publicdataset show that the performance of the proposed HRS with CRBMoutperforms current pattern recognition methods in SAR ATR, likePCA + SVM, LDA + SVM, and NMF + SVM.