多视角声图中水下小目标分类的深度神经网络方法研究

Small underwater objects classification in multi-view sonar images using the deep neural network

针对基于多视角声图的水下小目标分类问题,提出了一种深度神经网络多视分类方法。首先,提取声图的阴影区域,计算阴影部分的主轴斜率并匹配出与其相对应的仿真数据集。采用由这些对应仿真数据集训练的卷积神经网络分别对不同视角的待分类声图提取深度神经网络特征。将不同视角输出的特征向量组合起来,作为目标的特征向量,利用各个视角匹配的组合所对应的支持向量机对目标的特征向量进行预测。将分类器用于对湖、海试采集的多视角声图分类,平均正确率为93.33%,相比采用卷积神经网络、支持向量机的单视角分类方法,分别有不同程度的提升。

To solve the problem of small underwater objects classification based on multi-view sonar images, a deep neural network classification method with multi-view is proposed. Firstly, the shadow area of underwater objects in sonar images is extracted. The main axis slope of shadow area is calculated, which is used to match sonar images to the corresponding simulated dataset. The convolutional neural network trained by this simulated dataset is applied to extract deep neural network features from multi-view sonar images. The achieved feature vectors from sonar images of different views are combined as a feature vector of underwater object and predicted from support vector machine. The classifier is utilized to classify multi-view sonar images collected from lake and sea trials. The average classification accuracy can reach 93.33%. The performance is improved compared with the single-view classification method using convolutional neural network and support vector machine.