基于多波段深度神经网络的舰船目标识别

发布时间：2018-10-05 09:09

【摘要】：考虑多波段图像的融合识别可以扩展识别系统的应用范围,本文探索并设计了一种基于卷积神经网络的融合识别方法。该方法以AlexNet网络模型为基础,同时对可见光、中波红外和长波红外三波段图像进行特征提取;然后,利用互信息的方法对串联的三波段特征向量进行特征选择,依据重要性排序的方式选定固定长度的特征向量;最后,依据特征提取层级的不同,分别以早期融合、中期融合和后期融合3种融合方式来验证算法的有效性。采用自建的三波段舰船图像数据库进行了模型的训练和测试,共包含6类目标,5 000余张图像。实验结果显示,采用的3种融合识别方法中,中间层融合的识别准确率最高,达到84.5%,比早期融合和后期融合分别高5%和7%左右。另外,在本文的应用场景下,无论何种融合方式,其融合识别的准确率均明显高于其他单波段识别的准确率。
[Abstract]:Considering that multi-band image fusion recognition can extend the application range of recognition system, this paper explores and designs a fusion recognition method based on convolution neural network. The method is based on the AlexNet network model and extracts the features of the visible, medium and long wave infrared images simultaneously, and then uses mutual information to select the feature vectors in series. The feature vectors of fixed length are selected according to importance ranking. Finally, according to the different levels of feature extraction, the validity of the algorithm is verified by three fusion methods: early fusion, intermediate fusion and late fusion. The model is trained and tested by using the three-band ship image database, which contains more than 5,000 images of 6 kinds of targets. The experimental results show that the middle layer fusion has the highest recognition accuracy of 84.5, which is about 5% and 7% higher than that of early fusion and late fusion, respectively. In addition, the accuracy rate of fusion recognition is obviously higher than that of other single-band recognition in the application scenario of this paper.
【作者单位】：海军航空工程学院控制工程系;中国国防科技信息中心;91206部队;
【基金】：国家自然科学基金资助项目(No.61303192)
【分类号】：TP183;TP391.41

【相似文献】