| 本文已被：浏览 694次 下载 699次 |
码上扫一扫！ |
| 基于神经网络和机器学习的白羽肉鸡体重估测算法 |
|
庄超1,沈明霞1*,刘龙申1,姚文2,郑荷花3,王梦雨1
|
|
|
| (1.南京农业大学 工学院, 南京 210031;2.南京农业大学 动物科技学院, 南京 210031;3.新希望六合股份有限公司, 山东 青岛 266100) |
|
| 摘要: |
| 针对规模化肉鸡养殖生产中，传统肉鸡称重方法易造成应激问题，设计一种基于神经网络和机器视觉技术的非接触式肉鸡体重估测方法。应用深度相机采集白羽肉鸡的红外图像和深度信息，以目标识别算法YOLOv3和卷积网络分割算法FCN(Fully convolutional networks)为基础构建肉鸡区域提取模型，YOLOv3和FCN模型的查准率分别为98.1%和97.8%，查全率100%；结合肉鸡的深度信息，提取肉鸡投影面积等相关特征，构建多种回归算法，训练并调整优化肉鸡体重估测模型，ABR(Adaboost regressor)模型在测试集上达到最优的估测效果，该模型的决定系数R2为0.95，绝对误差为0.01~0.32 kg。本研究的非接触式的肉鸡体重估测模型能较好的预测肉鸡体重，为实际生产环境中肉鸡自动称重提供了技术支持。 |
| 关键词: 肉鸡体重 图像处理 神经网络 特征提取 回归模型 |
| DOI：10.11841/j.issn.1007-4333.2021.07.11 |
| 投稿时间：2020-10-10 |
| 基金项目:国家重点研发计划(2017YFD0701602-2) |
|
| Weight estimation model of breeding chickens based on neural network and machine learning |
|
ZHUANG Chao1,SHEN Mingxia1*,LIU Longshen1,YAO Wen2,ZHENG Hehua3,WANG Mengyu1
|
| (1.College of Engineering, Nanjing Agricultural University, Nanjing 210031, China;2.College of Animal Technology, Nanjing Agricultural University, Nanjing 210031, China;3.New Hope Liuhe Co., Ltd., Qingdao 266100, China) |
| Abstract: |
| Aiming at the problems of broiler weighing difficulty and easy to cause stress during the large-scale broiler breeding process, a non-contact broiler weight estimation method based on neural network and machine vision technology is adopted to study the change and monitoring of broiler weight. The results show that the infrared image and depth information of white feather broilers are collected by depth camera. Based on the target recognition algorithm YOLOv3 and the convolutional network segmentation algorithm FCN(Fully convolutional networks), a broiler region extraction model is constructed. The precisions of YOLOv3 and FCN models reach 98. 1% and 97. 8%, respectively, and the recall rate is 100%. Combining the depth images of broilers, extracting the relevant characteristics of broiler weight, constructing multiple regression model estimation algorithms, training and adjusting and optimizing the broiler weight estimation model, the Adaboost regressor model achieves the best estimation effect on the test set. The coefficient of determination is 0. 95, and the absolute error varies within the range of 0. 01-0. 32 kg. In conclusion, the estimation algorithm model proposed in this paper can better predict the weight of broilers, and provides technical support for the automatic weighing of broilers in the actual production environment. |
| Key words: chicken weight image processing neural network feature extraction regression model |
