Convolutional factor analysis model with application to radar automatic target recognition

Abstract

Convolutional networks, such as convolutional neural networks, are extensively applied in pattern recognition systems, where the convolution kernel plays an important role in performance improvement. Inspired by this, we integrate the convolution operation into statistical modeling in this paper and develop a novel probabilistic generative model called convolutional factor analysis (CFA), which is more applicable to statistical recognition with limited training data. And then, the CFA model is applied to radar automatic target recognition based on high-resolution range profile (HRRP), where sufficient training data are frequently unavailable due to the sampling rate limitation of real radar systems. As a dictionary learning method, the dimension of each dictionary atom in our CFA model is much lower than that in the traditional factor analysis (FA) model. Meanwhile, the model also makes it possible to capture the basic structures of observed data, thus requiring a fewer dictionary atoms to describe observations. Due to these two properties, the dictionary size of the CFA model is much smaller. Therefore, compared to the traditional FA model, the CFA model has the lower order of model complexity and can be learned better under a small amount of training data. In addition, owing to the conjugate property, the model parameters can be inferred via variational Bayesian (VB) algorithm, and the commutative law of convolution operation is also exploited to simplify the derivations of the posteriors. Experiments on synthetic and measured HRRP data show that basic structures of data can be represented by the dictionaries learned via our CFA model, and the better recognition performance can also be achieved by our method with small training data size.