渔业现代化

渔业现代化杂志

• 论文 • 上一篇    下一篇

基于ResNet的典型养殖鱼类识别方法研究

  1.  (1中国水产科学研究院渔业机械仪器研究所，上海 200092；
    2农业农村部渔业装备与工程技术重点实验室，上海 200092)
  • 出版日期:2022-06-20 发布日期:2022-10-28
  • 通讯作者: 刘世晶（1982—），男，硕士，副研究员，研究方向：渔业信息化、图像处理、模式识别和机器视觉。E-mail: liushijing@fmiri.ac.cn
  • 作者简介:涂雪滢（1992—），女，硕士，研究实习员，研究方向：机器视觉与图像处理、渔业信息化。E-mail: tuxueying@ fmiri.ac.cn
  • 基金资助:
    上海市科技兴农项目（沪农科推字(2019)第3-2号)

Study on the identification methods of typical cultured fish based on ResNet

  1. (1 Fishery Machinery and Instrument Research Institute, Chinese Academy of Fishery Sciences, Shanghai 200092, China;
    2 Key Laboratory of Fishery Equipment and Engineering, Ministry of Agriculture, Shanghai 200092, China)
  • Online:2022-06-20 Published:2022-10-28

PDF

摘要： 水下目标精准识别对指导养殖生产、辅助养殖决策具有十分重要的作用，而目标识别精度和运行效率是影响识别技术深入应用的关键问题。针对鱼类识别产业应用需求，以ResNet（Residual Neural Network）框架为核心，通过比较分析不同框架结构对鱼类识别精度和效果的影响，确定适用于典型养殖鱼类识别的ResNet网络结构形式。首先，采用多相机同步采样方式，获取不同姿态鱼类图像，满足高柔性、多姿态的运动目标样本集构建需要；其次，为了提升样本对不同背景的适应能力，选取具有不同背景的目标鱼类图像，丰富图像样本集；然后，以典型的ResNet18、ResNet34、ResNet50框架结构为比较模型，分析不同结构在识别效率和识别精度方面的整体效果。结果显示，ResNet50识别精度最高，达到95.47%，ResNet34次之，达到95.03%，但ResNet50识别效率比ResNet34降低20.43%，综合考虑识别精度和识别效率，ResNet34更加适用于大样本量鱼类图像的识别分类。

关键词: 鱼类识别, 深度学习, 分类识别, ResNet模型

Abstract: Accurate identification of underwater targets plays a very important role in guiding aquaculture production and assisting aquaculture decision-making, and target identification accuracy and operational efficiency are the key problems affecting the in-depth application of recognition technology. In view of the application requirements of the fish identification industry, this paper takes the ResNet (Residual Neural Network) framework as the core, compares and analyses the impact of different framework structures on fish identification accuracy and effect, and determines the ResNet structure form suitable for typical breeding fish identification. Firstly, multi-camera synchronous sampling to obtain different fish images to meet the needs of a highly flexible and multi-attitude motion target sample set. Secondly, in order to improve the adaptability of samples to different backgrounds, selecting target fish images with different backgrounds to enrich the image sample set. Then, comparing the typical ResNet 18, ResNet 34, and ResNet 50 framework to analyze the overall effect of different structures on identification efficiency and identification accuracy. The rest results show that the ResNet 50 has the highest recognition accuracy at 95.47%, followed by ResNet 34 and 95.03%, but the ResNet 50 recognition efficiency is 20.43% lower than ResNet 34. Considering the recognition accuracy and recognition efficiency, ResNet 34 is more suitable for the recognition classification of fish images with a large sample size.

Key words: fish identification, deep learning, classification and identification, ResNet model;