基于卷积自动编码器的飞机目标识别方法  

An Airplane Target Recognition Method Based on Convolutional Auto-encoder

作　　者：史珊珊 史鹤欢 SHI Shan-shan;SHI He-huan(School of Electronic Engineering,Xi’an Aeronautical University,Xi’an 710077,China;Aeronautics and Astronautics Engineering college,Air Force Engineering University,Xi’an 710038,China)

机构地区：[1]西安航空学院电子工程学院,西安710077 [2]空军工程大学航空航天工程学院,西安710038

出　　处：《火力与指挥控制》2019年第8期23-28,共6页Fire Control & Command Control

基　　金：国家自然科学基金面上项目（61379104）

摘　　要：飞机目标标签数据不足,使传统的机器学习算法训练效率不足。为提升训练效率提高飞机目标识别率,提出一种由卷积自动编码器(Convolutional Auto-Encoder,CAE)、哈希变换及直方图统计组成的简单多层特征提取模型。该模型利用CAE非监督训练一组卷积滤波器,与输入数据卷积提取特征;并再次利用CAE训练卷积滤波器集合,提取卷积特征;对所得到的卷积特征进行哈希变换和直方图统计;用支持向量机识别分类。实验对飞机目标取得了较高的识别率,表明特征提取模型具有很强的鲁棒性。Traditional machine learning algorithms is inefficient in training because of the lack of military airplane label data. A simple multi-layer feature extracting model that improves training efficiency and recognition rate is proposed which consist of Convolutional Auto-Encoder(CAE),binary and block-wise histograms. First,a set of filters are used by CAE and the features are convolutional results of the filters and input data. Then,the CAE is used to train the filters and the features are extracted again. Next,we binarize the previous features and compute the histograms. Finally,Support Vector Machine(SVM) is used to recognize and classify the targets. Experiments achieves a higher recognition rate for military airplane and demonstrates the feature extracting model has a better robustness for large variations.

关 键 词：卷积自动编码器 哈希变换 直方图统计 军用飞机 特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]