欢迎您访问《智慧农业(中英文)》官方网站! English

Smart Agriculture ›› 2022, Vol. 4 ›› Issue (2): 163-173.doi: 10.12133/j.smartag.SA202201012

• 信息感知与获取 • 上一篇    下一篇

基于改进Mask R-CNN模型的工厂化养蚕蚕体识别与计数

何锐敏1(), 郑可锋2, 尉钦洋1, 张小斌2, 张俊1, 朱怡航2, 赵懿滢2, 顾清2()   

  1. 1.嵊州陌桑高科股份有限公司,浙江 绍兴 312400
    2.浙江省农业科学院数字农业研究所,浙江 杭州 310021
  • 收稿日期:2021-11-02 出版日期:2022-06-30
  • 基金资助:
    浙江省重点研发计划项目(2019C02001)
  • 作者简介:何锐敏(1964-),女,学士,研究方向为工厂化养蚕。E-mail:agri_intel@163.com
  • 通信作者: 顾 清(1986-),男,博士,副研究员,研究方向为农业信息技术。E-mail:guq@zaas.ac.cn

Identification and Counting of Silkworms in Factory Farm Using Improved Mask R-CNN Model

HE Ruimin1(), ZHENG Kefeng2, WEI Qinyang1, ZHANG Xiaobin2, ZHANG Jun1, ZHU Yihang2, ZHAO Yiying2, GU Qing2()   

  1. 1.Shengzhou Mosang High-tech Co. , Ltd. , Shaoxing 312400, China
    2.Institute of Digital Agriculture, Zhejiang Academy of Agricultural Sciences, Hangzhou 310021, China
  • Received:2021-11-02 Online:2022-06-30
  • Supported by:
    Key Research and Development Plan Project of Zhejiang Province (2019C02001)

摘要:

精准饲喂是全龄人工饲料工厂化养蚕节本增效的核心技术之一,家蚕自动化识别与计数是实现精准饲喂的关键环节。本研究基于机器视觉系统获取工厂化养蚕过程中蚕在4龄和5龄期的数字图像,利用改进深度学习模型Mask R-CNN检测蚕体和残余饲料。通过在Mask R-CNN模型框架中加入像素重加权策略和边界框细调策略,从噪声数据中训练一个鲁棒性更好的目标检测模型,实现模型性能的优化,提高对蚕体和饲料边界的检测和分割能力。改进Mask R-CNN模型对蚕的检测和分割交并比阈值为0.5时的平均精度(Average Precision at IoU=0.5,AP50)分别为0.790和0.795,识别准确率为96.83%;对残余饲料的检测和分割AP50分别为0.641和0.653,识别准确率为87.71%。模型部署在NVIDIA Jetson AGX Xavier开发板上,单张图像平均检测时间为1.32 s,最长检测时间为2.05 s,运算速度可以满足养蚕盒单元在生产线上移动实时检测的要求。该研究为工厂化养蚕精准饲喂信息系统和投喂装置的研发提供了核心算法,可提高人工饲料的利用率,提升工厂化养蚕生产管理水平。

关键词: 家蚕, 人工饲料, 精准饲喂, 机器视觉, 深度学习, mask R-CNN, 噪声数据

Abstract:

Factory-like rearing of silkworm (Bombyx mori) using artificial diet for all instars is a brand-new rearing mode of silkworm. Accurate feeding is one of the core technologies to save cost and increase efficiency in factory silkworm rearing. Automatic identification and counting of silkworm play a key role to realize accurate feeding. In this study, a machine vision system was used to obtain digital images of silkworms during main instars, and an improved Mask R-CNN model was proposed to detect the silkworms and residual artificial diet. The original Mask R-CNN was improved using the noise data of annotations by adding a pixel reweighting strategy and a bounding box fine-tuning strategy to the model frame. A more robust model was trained to improve the detection and segmentation abilities of silkworm and residual feed. Three different data augmentation methods were used to expand the training dataset. The influences of silkworm instars, data augmentation, and the overlap between silkworms on the model performance were evaluated. Then the improved Mask R-CNN was used to detect silkworms and residual feed. The AP50 (Average Precision at IoU=0.5) of the model for silkworm detection and segmentation were 0.790 and 0.795, respectively, and the detection accuracy was 96.83%. The detection and segmentation AP50 of residual feed were 0.641 and 0.653, respectively, and the detection accuracy was 87.71%. The model was deployed on the NVIDIA Jetson AGX Xavier development board with an average detection time of 1.32 s and a maximum detection time of 2.05 s for a image. The computational speed of the improved Mask R-CNN can meet the requirement of real-time detection of the moving unit of the silkworm box on the production line. The model trained by the fifth instar data showed a better performance on test data than the fourth instar model. The brightness enhancement method had the greatest contribution to the model performance as compared to the other data augmentation methods. The overlap between silkworms also negatively affected the performance of the model. This study can provide a core algorithm for the research and development of the accurate feeding information system and feeding device for factory silkworm rearing, which can improve the utilization rate of artificial diet and improve the production and management level of factory silkworm rearing.

Key words: silkworm, artificial diet, accurate feeding, machine vision, deep learning, mask R-CNN, noise data

中图分类号: