An infinite classification RBM model for radar HRRP recognition

High resolution range profile (HRRP) recognition is a widely used technique in the radar automatic target recognition community. However, the recognition performance often suffers from aspect sensitivity and the unsuitable features designed in practical applications. Aiming at dealing with these problems, we propose a new type of stochastic neural network model named Infinite Classification Restricted Boltzmann Machine (RBM) in this paper, which originates from the Infinite RBM and the Classification RBM. Different from most conventional methods using separate models for each aspect frame, the proposed model can jointly model HRRPs from different aspects with a unified stochastic model, which benefits from its power of learning complex distribution of data. Besides that, our model can adaptively learn suitable features according to specific recognition tasks. Owing to these two properties, the proposed model learns discriminative features more efficiently with relative smaller training data than other traditional recognition techniques. Experiment results using simulated HRRP data indicate that our model has more generalization power and better recognition performance than the Classification RBM.