计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (17): 236-242.DOI: 10.3778/j.issn.1002-8331.1905-0317

• 工程与应用 • 上一篇    下一篇

不平衡数据集下的水下目标快速识别方法

刘有用,张江梅,王坤朋,冯兴华,杨秀洪   

  1. 西南科技大学 信息工程学院,四川 绵阳 621010
  • 出版日期:2020-09-01 发布日期:2020-08-31

Fast Underwater Target Recognition with Unbalanced Data Set

LIU Youyong, ZHANG Jiangmei, WANG Kunpeng, FENG Xinghua, YANG Xiuhong   

  1. School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
  • Online:2020-09-01 Published:2020-08-31

摘要:

水下目标的准确识别是水下机器人实现抓取、捕捞等安全作业的前提,针对水下图像质量差、样本数量少及类不平衡而导致目标识别精确度低的问题,提出了一种基于生成对抗网络(Generative Adversarial Networks, GAN)的水下目标快速识别算法。利用GAN理论搭建了深度卷积神经网络的水下图像生成模型,通过生成器与判别器的零和博弈生成特定水下目标图像;设计生成目标的中心坐标计算函数和边界融合函数,将生成目标与背景图像融合后训练水下目标识别模型。实验结果表明,所提方法能够显著提高水下目标识别精确度,对实现水下目标准确抓取、促进水下作业及海洋资源的开发具有重要意义。

关键词: 生成对抗网络(GAN), 图像生成, 深度神经网络, 水下目标识别

Abstract:

Accurate recognition of underwater targets is the prerequisite for the realization of underwater robot capture and fishing operations. Aim at the problem that the poor quality of underwater images, small sample size or class imbalance lead to low accuracy of target recognition, a fast recognition method based on Generative Adversarial Networks(GAN) is proposed.The underwater image generation model of deep convolutional neural network is built by using GAN theory, and the specific underwater target image is generated by the zero-sum game between the generator and discriminator. The central coordinate calculation function and boundary fusion function of the generated target are designed. The generated target and the background image are fused and then train underwater target recognition model.Experimental results show that the method proposed in this paper can significantly improve the accuracy of underwater target identification, which is of great significance for the realization of accurate capture of underwater targets, the promotion of underwater operations and the development of ocean resources.

Key words: Generative Adversarial Networks(GAN), image generation, deep neural network, underwater target recognition