基于模糊不变卷积神经网络的遥感飞机识别  

Remote Sensing Aircraft Recognition Based on Blur-Invariant Convolutional Neural Network

作　　者：刘坤[1] 苏彤 王典 Liu Kun;Su Tong;Wang Dian(College of Information Engineering,Shanghai Maritime University,Shanghai 200135,China)

机构地区：[1]上海海事大学信息工程学院,上海200135

出　　处：《激光与光电子学进展》2018年第8期387-394,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61271446);航空科学基金(2013ZC15005)

摘　　要：提出了一种基于模糊不变卷积神经网络(BICNN)模型的目标识别方法。与传统卷积神经网络(CNN)模型不同,BICNN引入了一种新的模糊不变层。BICNN通过增加模糊不变约束项及正则化来优化模糊不变目标函数并进行训练;通过减小模糊不变目标函数值,使得训练样本在模糊前后的特征映射一致,最终实现模糊不变性。测试结果表明,BICNN解决了模糊造成的识别率低的问题,增大了运动模糊图像的识别率。A method of target recognition based on the blur-invariant convolutional neural network（BICNN）model is proposed.The BICNN model introduces a new blur-invariant layer,which is different from the traditional convolutional neural network（CNN）models.BICNN is trained by the adding of the blur-invariant constraint term and the regularization to optimize a blur-invariant objective function.The value of the fuzzy invariant objective function is reduced to make the training samples consistent with the feature maps before and after the blurring,and thus the blur invariance is achieved finally.The test results show that BICNN can solve the problem of a low recognition rate caused by blur and improve the recognition rate of the motion blurred images.

关 键 词：光计算 深度学习 卷积神经网络 目标识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]