Lightweight CNN for Radar HRRP Recognition Using NAS-Based Pruning and Multi-Knowledge Distillation

Deep convolutional neural network (CNN) has been widely studied in radar target high resolution range profile (HRRP) recognition. However, the CNN with deep structure requires high storage and computational capabilities, thus restricting its applications with limited resources. In this paper, we propose a differentiable neural architecture search (NAS)-based channel pruning and multi-knowledge distillation (MKD) to prune and fine-tune the CNN for lightweight. In the pruning stage, NAS method is used to automatically find the good-performance pruning structure by optimizing channel preserve scores. After the search process, these scores are used for global pruning to derive the pruned model. In the fine-tuning stage, MKD method combine the physical scattering knowledge with the category knowledge from pretrained model to restore the pruned model performance. Experimental results on VGG-16 with HRRP dataset inversed from the MSTAR show that the recognition accuracy of the proposed method decreases 1.52% by using 5% model parameters.

I. Introduction

High resolution range profile (HRRP) is the sum of the target's scattering center echo along the sight of radar, which contains the message of the target structure and texture [1]. Due to the easy accessibility and processing of HRRP, it has been widely used in the scene of real-time radar target recognition.