检索
E-mail
RSS

作者投稿 主编审稿 专家审稿 远程编辑

热带海洋学报

• • 上一篇    下一篇

融合实例分割的双目视觉鱼类尺寸原位测量方法

孙乾泰1,2，孙亮1，雷平顺1，林先智4，陈嘉男1，张钰润1,2，周燕1,2,3，刘育梁1,2，刘胜4，王新伟1,2,3

  

  1. 1. 中国科学院半导体研究所光电系统实验室，北京，100083; 

    2. 中国科学院大学材料与光电研究中心，北京，100049；

    3. 中国科学院大学电子电气与通信工程学院，北京，100049;

    4. 中国科学院南海海洋研究所，广州，510301




  • 收稿日期:2026-04-27 修回日期:2026-07-14 接受日期:2026-07-20
  • 通讯作者: 孙亮
  • 基金资助:
    国家科技基础资源调查专项(2022FY100605); 国家重点研发计划(2024YFC2815203); 国家自然科学基金项目(42276197)

A Binocular Vision Method for In Situ Fish Size Measurement Integrating Instance Segmentation

Sun Qiantai1,2, Sun Liang1, Lei Pingshun1, Lin Xianzhi4, Chen Jianan1,Zhang Yurun1,2,Zhou Yan1,2,3, Liu Yuliang1,2, Liu Sheng4, Wang Xinwei1,2,3    

  1. 1 Optoelectronic System Laboratory, Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China;

    2 School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China;

    3 Center of Materials Science and Optoelectronics Engineering, University of Chinese Academy of Sciences, Beijing 100049, China;

    4South China Sea Institute of Oceanology, Chinese Academy of Sciences, Guangzhou 510301, China




  • Received:2026-04-27 Revised:2026-07-14 Accepted:2026-07-20
  • Supported by:

    Science & Technology Fundamental Resources Investigation Program (2022FY100605); National Key Research and Development Program of China (2024YFC2815203); the National Science Foundation of China (42276197)

PDF (PC)

赞

摘要： 针对海洋生态监测中鱼类尺寸原位、自动测量的需求，本文提出了一种融合实例分割先验信息的双目视觉鱼类尺寸原位测量方法，旨在解决水下复杂环境对传统双目视觉立体匹配精度的干扰问题。该方法首先采用YOLOv8-Seg实例分割网络提取精确的鱼体掩码，并将该掩码作为空间先验引导半全局块匹配算法，仅于目标区域内计算视差，在有效抑制背景噪声、降低误匹配的同时提升匹配速度，进而对目标区域深度峰值进行提取与分析，采用视差频率峰值法获取鱼体代表深度，最终结合相机标定参数将掩码像素尺度映射至物理空间，从而实现鱼体尺寸的高精度估算。基于上述方法，本文研制了水下双目立体相机“水睛-Trap”。受控水池实验结果表明，目标尺寸测量的平均相对误差为5.59%，变异系数为6.68%，测量精度达厘米级；同时得益于分割先验信息的引入，尺寸测量速度提升约3倍，能够满足实时性观测需求。在西沙海试中，“水睛-Trap”在不同深度和光照环境下运行稳定，测量结果表现出较高一致性（变异系数为3.34%）。研究结果表明，本文所提出的“分割先验+立体匹配”的策略在测量精度和效率之间取得了良好平衡，可为鱼类的“自动化、数字化、精细化“原位监测提供有效的技术支撑。

关键词: 光学原位监测, 双目立体视觉, 实例分割, 鱼类尺寸, 水下三维重建

Abstract: To meet the demand for automated in situ fish size measurement in marine ecological monitoring, this study proposes a binocular stereo vision method incorporating instance segmentation priors, aiming to mitigate the interference of complex underwater environments with the accuracy of conventional stereo matching. The proposed method first employs the YOLOv8-Seg instance segmentation network to extract accurate fish masks, which are then used as spatial priors to guide the Semi-Global Block Matching algorithm. In this way, disparity computation is restricted to the target region, effectively suppressing background noise, reducing mismatches, and improving processing speed. The depth peak within the target region is subsequently extracted and analyzed, and the representative fish depth is determined using a disparity-frequency peak-based method. Finally, camera calibration parameters are incorporated to map the pixel-scale mask to physical space, enabling high-precision estimation of fish size. Based on this methodology, an underwater stereo camera system named “Shuijing-Trap” was developed. Controlled tank experiments showed that the proposed method achieved a mean relative error of 5.59% and a coefficient of variation of 6.68% in target size measurement, corresponding to centimeter-level accuracy. Moreover, owing to the introduction of segmentation priors, the measurement speed was improved by approximately three times, satisfying the requirements for real-time observation. In subsequent sea trials conducted in the Xisha Islands, the “Shuijing-Trap” system operated stably under varying depths and illumination conditions, and the measurement results showed high consistency, with a coefficient of variation of 3.34%. These results indicate that the proposed “segmentation prior + stereo matching” strategy achieves a favorable balance between measurement accuracy and efficiency, providing effective technical support for automated, digital, and fine-scale in situ monitoring of fish populations.

Key words: optical in situ monitoring, stereo vision, instance segmentation, fish size measurement, underwater 3D reconstruction