Micro-motion Based Target Recognition Using Regularized DCNN

Micro-motion is an unique status of moving targets, playing an important role in detecting and recognizing space targets. Convolutional neural networks (CNNs) has been widely applied for target recognition based on micro-motions, but the extracted features are still sensitive to environmental noise, thus it is of vital importance to improve CNNs' robustness to micro-motion targets under noisy environments. This paper takes ballistic missiles as an example, establishes and thus analyzes the echo signal model of cone warhead under three micro-motion conditions: rotation, coning and vibration. Based on the simulated radar echo signals under hypothetical war scenario, a dataset of time-frequency spectrums is established, consisting of twelve different types of ballistic targets which include warheads as well as decoys. This paper applys a regularization method to DCNNs, by minimizing feature variations caused by gaussian noise, notably improves the recognition accuracy of micro-motion targets as well as its robustness to gaussian noise, compared with conventional CNN methods.