A research on underwater target recognition neural network for small samples

面向小样本数据的水下目标识别神经网络深层化研究

School of Marine Science and Technology, Northwestern Polytechnical University, Xi′an 710072, China

Received: 20 May 2021

Abstract

In the face of the challenges in the field of marine engineering applications in the new era, the goal of automation, high efficiency and accuracy can be achieved by using deep learning-based neural networks in hydroacoustic engineering. However, in the face of objective problems such as the lack of underwater target samples, the complex underwater sound environment, and the poor sample signal-to-noise ratio, the deep learning also becomes less sensitive due to its own limitations. In this paper, by constructing a variety of target feature extraction methods and a deep neural network model, we obtain the target recognition rate network prediction value after matched different target feature extraction with neural network model. Through comparing experimental results, a new idea of solving small sample target identification through deep neural network deep design is proposed.

摘要

在面对新时期海洋工程应用领域的挑战时, 可以通过利用基于深度学习的神经网络在水声工程中的实现, 来达成自动化、高效性、准确性的目标。然而在面对水下目标样本匮乏、水下声环境复杂、样本信噪比差等客观问题时, 深度学习也会因其自身的局限性而变得不那么灵敏。针对小样本问题, 通过构建多种目标特征提取法和深层深度神经网络模型, 得到了不同目标特征提取与网络模型匹配后的目标识别率与网络预测值, 并通过比对实验结果, 提出了通过深层神经网络深层化设计解决小样本目标识别的新思路。