用于水声目标特征学习与识别的混合正则化深度置信网络

发布时间：2018-08-29 19:02

【摘要】：针对获取水声目标有类标样本困难且代价高昂的水声目标小样本识别问题,提出了基于混合正则化深度置信网络(hybrid regularization deep belief network,HR-DBN)的水声目标深度特征学习及识别方法。该方法首先提出了混合2种正则化策略的深度置信网络进行水声目标深度特征学习。第一种正则化策略是利用最大互信息组正则化项修正目标函数,提高隐含层的稀疏度;第二种正则化策略是利用大量无类标样本获得有关水声目标的普遍特性的描述和先验知识,引导特征学习。最后利用少量有类标样本对网络进行全局优化,构建识别系统,提高水声目标识别正确率。利用2类实测舰船辐射噪声数据进行验证实验,实验结果表明,提出的方法可以提取描述水声目标的深度特征,提高水声目标识别正确率。
[Abstract]:Aiming at the problem of small sample recognition of underwater acoustic targets which is difficult and costly to obtain underwater acoustic targets, a method of depth feature learning and recognition of underwater acoustic targets based on mixed regularized depth confidence network (hybrid regularization deep belief network,HR-DBN) is proposed. In this method, a depth confidence network with two regularization strategies is proposed to study the depth characteristics of underwater acoustic targets. The first regularization strategy is to modify the objective function by using the maximal mutual information set regularization term to improve the sparse degree of the hidden layer. The second regularization strategy is to obtain the description and prior knowledge of the general characteristics of underwater acoustic targets by using a large number of unlabeled samples to guide feature learning. Finally, a small number of labeled samples are used to optimize the network globally, and a recognition system is constructed to improve the accuracy of underwater acoustic target recognition. The experimental results show that the proposed method can extract the depth characteristics of underwater acoustic targets and improve the accuracy of underwater acoustic target recognition.
【作者单位】：西北工业大学航海学院;
【基金】：盲信号处理重点实验室基金资助
【分类号】：TB56

【相似文献】