水中目标特征增强及深度半监督分类方法研究

The technique of passive underwater target recognition has promising applications prospect in the areas of military detection, ocean search and rescue, resource exploration,etc. The scarcity of labeled samples and the insufficient of robustness in real environment of current researches are the key problems, which have restricted the technique towards to practical. This project will bring forward a novel deep semi-supervised recognition method which accords with both the characteristics of noise emitted by underwater targets and the practical requirements. Firstly, the relations between the feature distortion, classification performance and the above two problems will be studied thoroughly to reveal the regulations of influence. Then, the feature extraction and enhancement methods based on structural sparse decomposition theory, the channel impact suppression method based on cepstrum and factor analysis, a sliding average window method based on logarithmic power spectrum will be suggested to enhance the acoustic features. Finally the deep neural network and the Dirichlet process are combined to realize the semi-supervised classification. This novel method combines the deep learning network, sparse feature extraction, feature learning, feature enhancement, and semi-supervised classification into one system which can realize the online learning of parameters and the auto modeling of hidden class information. This study can enhance the adaption ability of the unsupervised classification algorithm and will provide important foundation for the engineering applications of the technique.

水中目标被动识别研究在军事探测、海洋搜救、资源勘探等领域具有广阔的应用前景。标记样本欠缺和实际环境中鲁棒性不强是制肘该技术走向实用的关键问题。本项目提出一种新的符合水下目标辐射噪声特性和实用需求的深度半监督分类识别方法。首先对前述两方面问题与特征畸变和识别性能之间的关系进行深入研究，揭示影响规律。在此基础上，分别提出基于结构化稀疏分解的特征提取与增强方法、基于倒谱处理和因子分析模型的通道影响抑制方法和一种对数功率谱滑动窗平均方法，用于特征增强。最后，融合深度神经网络和Dirichlet过程实现半监督分类识别。这种新的识别方法以深度学习网络为载体，将信号稀疏去噪、结构化稀疏特征提取、特征增强、特征学习和半监督分类决策融为一个整体，能实现关键参数的在线学习和隐藏类别信息的自动建模，从而提高分类识别方法的自适应能力，为该技术走向工程应用提供重要的理论基础。

水中目标被动识别研究在军事探测、海洋搜救、资源勘探等领域具有广阔的应用前景。实际应用中存在标记样本欠缺和鲁棒性不强等关键难题。本项目首先对前述两方面问题与特征畸变和识别性能之间的关系进行深入研究，揭示影响规律。在此基础上，分别提出基于结构化稀疏分解的特征提取与增强方法、基于倒谱处理和因子分析模型的通道影响抑制方法和一种对数功率谱滑动窗平均方法，用于特征增强。最后，融合深度神经网络和Dirichlet过程提出一种新的符合水下目标辐射噪声特性和实用需求的深度半监督分类识别方法。该方法以深度学习网络为载体，将信号稀疏去噪、结构化稀疏特征提取、特征增强、特征学习和半监督分类决策融为一个整体，能实现关键参数的在线优化和隐藏类别信息的自动建模，从而提高分类识别方法的自适应能力。项目通过实测数据对提出的方法进行了有效性验证。