基于深度卷积特征的水下目标智能识别方法

Underwater Target Recognition Method Based on Deep Convolution Feature

无人水下航行器(UUV)由于本身的便利性和自主性在水下探测方面相比传统探测具有很大优势,对UUV水下目标智能识别方法的研究具有重要意义。针对水下环境的高噪声、低对比度的特点,文中首先使用中值滤波和局部增强处理进行水下图像预处理,基于水下图像的小样本特点,提出借鉴牛津大学视觉几何组网络(VGGNet)的逐层递增卷积层思想,利用深度卷积神经网络(DCNN)设计水下智能识别框架并利用大数据集Cifar-10进行一次训练,以学习图像通用特征;同时使用迁移学习和数据增强技术进行二次学习,以学习水下目标特有特征,解决水下数据集不足的情况,防止过拟合。通过仿真试验进行识别效果验证,仿真结果表明,在特定测试集下提出的水下目标智能识别方法在识别效果与自动化程度方面相比传统识别算法具有明显优势。

With the help of artificial intelligence technology, the intelligent recognition technology for unmanned un- dersea vehicle(UUV) to detect underwater target is presented. According to the characteristics of high noise and low contrast in underwater environment, median filtering and local enhancement are adopted to preprocess underwater im- ages. Based on the characteristics of small sample of underwater images, the idea of incremental convolution layer by layer is proposed drawing from the visual geometry group net(VGGNet), and an underwater intelligent recognition framework is designed by using deep convolution neural network (DCNN) for training on the large data set Cifar-10 and learning general features of the image. In addition, migration learning and data enhancement technology are used for transferring knowledge to learn the unique features of underwater targets, so as to solve the problem of insufficient un- derwater data sets and to prevent over-fitting. Simulation results show that the proposed underwater target intelligent recognition technology has obvious advantages over the traditional algorithm in recognition effect and automation de- gree on specific test set.