用于弹载线阵红外与激光扫描成像引信的轻量化卷积神经网络目标识别方法  

A Lightweight Convolutional Neural Network Target Recognition Method for Missile-borne Fuse Linear Array Infrared and Lidar-scanning Image

作　　者：武军安 郭锐[1] 刘荣忠[1] 柯尊贵[2] WU Jun'an;GUO Rui;LIU Rongzhong;KE Zungui(ZNDY Ministerial Key Laboratory,Nanjing University of Science and Technology,Nanjing 210094,China;NORLA Institute of Technical Physics,Chengdu 610041,China)

机构地区：[1]南京理工大学智能弹药技术国防重点学科实验室,江苏南京210094 [2]西南技术物理研究所,四川成都610041

出　　处：《弹道学报》2021年第3期89-96,共8页Journal of Ballistics

基　　金：装备重点预研(301070201)。

摘　　要：为了进一步提高末敏弹在复杂战场环境下对地面装甲目标的识别概率,提出了一种基于轻量化卷积神经网络的红外图像与距离像复合探测识别方法。网络设计考虑了弹载环境对实时性的要求,将网络划分为特征提取、特征融合2个阶段。在特征提取阶段,对不同源的图像进行了分布式卷积,提高了网络的并行性,降低了网络参数量与计算量;为了弥补分布式卷积带来的特征损失,将距离像与红外图像的融合图像也一并作为网络输入;在特征融合阶段利用深度可分离卷积实现了进一步的轻量化设计。通过仿真缩比实验环境获得的数据集进行实验验证,实验结果表明,网络具有较小计算复杂度的同时能够对复杂背景环境下的装甲目标进行有效识别。In order to further improve the recognition probability of terminal sensitive projectile on ground armored targets in a complex battlefield environment,a composite detection and recognition method of infrared image and distance image based on lightweight convolutional neural network was proposed.The real-time requirement of the missile-borne environment was considered in the network design,and the network was divided into two stages:feature extraction and feature fusion.In the feature extraction stage,distributed convolution was performed on images from different sources,which improved the parallelism of the network and decreased the quantity of network parameters and computational complexity.At the same time,in order to compensate for the loss of features caused by distributed convolution,it was proposed that the fusion image of the distance image and the infrared image was also used as the network input;in the feature fusion stage,the depthwise separable convolution was used to achieve a further lightweight design.The experimental verification was carried out by using the data set obtained in the simulated scaled-down test.The results show that the proposed network can effectively identify armored targets in a complex background environment while having a small computational complexity.

关 键 词：末敏弹 线阵扫描 红外探测器 激光雷达 卷积神经网络 

分 类 号：TN958[电子电信—信号与信息处理]