基于深层卷积神经网络的肉兔图像分割与体质量估测  

Meat Rabbit Image Segmentation and Weight Estimation Model Based on Deep Convolution Neural Network

作　　者：段恩泽[1] 方鹏[1] 王红英[1] 金楠[1] DUAN Enze;FANG Peng;WANG Hongying;JIN Nan(College of Engineering, China Agricultural University, Beijing 100083, China)

机构地区：[1]中国农业大学工学院,北京100083

出　　处：《农业机械学报》2021年第6期259-267,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：财政部和农业农村部:国家现代农业产业技术体系项目(CARS-43-D-2)。

摘　　要：针对肉兔饲养管理过程中人工称量造成的应激、体质量信息采集困难等问题,提出了一种基于深层卷积神经网络的肉兔图像分割与体质量估测方法,实现了肉兔养殖管理中的无接触式称量。构建基于Mask RCNN的肉兔图像分割网络,以残差网络ResNet101作为主干网络,利用COCO数据集进行迁移学习以提高训练效率,获取围栏中不受限制的肉兔图像分割结果。提取每个样本掩膜的像素面积,通过引入弯曲度和体长两个特征参数来修正每个样本与对应体质量之间的权重关系。以投影面积、弯曲度、体长和日龄为输入参数,以肉兔体质量为输出参数,构建6神经元的体质量估测神经网络。分别测试肉兔图像分割网络和体质量估测神经网络,结果表明,肉兔图像分割网络在交并比(IoU)为0.5∶0.95时分类准确率为94.5%,对像素分割的精确度为95.1%。体质量估测神经网络的拟合相关系数R为0.99391,验证集均方误差为0.0336,预测体质量和实际体质量平均相差123 g。本文方法对不同日龄和不同姿态下肉兔的预测效果良好。In order to solve the problems in the process of feeding and management of meat rabbits,such as stress caused by manual weighing and difficulty in process of collecting weight information,a method of image segmentation and weight estimation based on deep convolution neural network was proposed,which can realize the contactless weighing of rabbits.A rabbit instance segmentation network based on Mask RCNN was constructed.The residual network ResNet101 was used as the backbone network,and COCO dataset was used for migration learning to improve the training efficiency and obtain the segmentation results of unrestricted meat rabbits in the fence.Then the pixel area of each sample mask was extracted,and curvature and body length were introduced to modify the weight relationship between each sample and the corresponding weight.Projection area,curvature,body length and age as input parameters and body weight as output parameters,a six neuron weight estimation neural network was constructed to test the rabbit instance segmentation network and weight estimation neural network,the results showed that when IoU was 0.5∶0.95,the classification accuracy of rabbit segmentation network was 94.5%,and the pixel segmentation accuracy was 95.1%.The fitting correlation coefficient R of weight estimation neural network was 0.99391,MSE was 0.0336,and the mean weight error was 123 g.The model had a good prediction effect on meat rabbits of different ages and different postures.

关 键 词：肉兔 深度学习 图像分割 体质量估测 

分 类 号：S829.1[农业科学—畜牧学] TP391.41[农业科学—畜牧兽医]