推荐星级:
  • 1
  • 2
  • 3
  • 4
  • 5

基于深度学习的群猪图像实例分割方法

更新时间:2019-12-31 09:43:41 大小:2M 上传用户:songhuahua查看TA发布的资源 标签:深度学习 下载积分:1分 评价赚积分 (如何评价?) 打赏 收藏 评论(0) 举报

资料介绍

群养饲喂模式下猪群有聚集在一起的习性,特别是躺卧时,当使用机器视觉跟踪监测猪只时,图像中存在猪体粘连,导致分割困难,成为实现群猪视觉追踪和监测的瓶颈。根据实例分割原理,把猪群中的猪只看作一个实例,在深度卷积神经网络基础上建立Pig Net网络,对群猪图像尤其是对粘连猪体进行实例分割,实现独立猪体的分辨和定位。Pig Net网络采用44层卷积层作为主干网络,经区域候选网络(Region proposal networks,RPN)提取感兴趣区域(ROI),并和主干网络前向传播的特征图共享给感兴趣区域对齐层(Region of interest align,ROIAlign),分支通过双线性插值计算目标空间,三分支并行输出ROI目标的类别、回归框和掩模。Mask分支采用平均二值交叉熵损失函数计算独立猪体的目标掩模损失。连续28 d采集6头9. 6 kg左右大白仔猪图像,抽取前7 d内各不同时段、不同行为模式群养猪图像2 500幅作为训练集和验证集,训练集和验证集的比例为4∶1。结果表明,Pig Net网络模型在训练集上总分割准确率达86. 15%,在验证集上准确率达85. 40%。本文算法对不同形态、粘连严重的群猪图像能够准确分割出独立的猪个体目标。将本文算法与Mask R-CNN模型及其改进模型进行对比,准确率比Mask RCNN模型高11. 40个百分点,单幅图像处理时间为2. 12 s,比Mask R-CNN模型短30 ms。


部分文件列表

文件名 大小
基于深度学习的群猪图像实例分割方法.pdf 2M

【关注B站账户领20积分】

部分页面预览

(完整内容请下载后查看)
2 0 1 9  
4
农 业 机 械 学 报  
50  
4
卷 第 期  
doi: 10. 6041 /j. issn. 1000-1298. 2019. 04. 020  
基于深群猪像实法  
12  
1
12  
23  
4
12  
高 云  
郭继亮 黎 煊  
刚  
童 宇  
( 1.  
中农业大学工学院 武汉  
430070; 2.  
同创新中心 武汉  
430070;  
3.  
中农业大学技学院学院 武汉  
430070; 4.  
中农业大学学院 武汉  
430070)  
: , , ,  
摘要 模式使用机视觉跟踪时 图存在猪  
, , ,  
导致困难 视觉瓶颈 根据中的作一实  
积神基础上建立  
PigNet  
体进行实分  
( Region proposal networksRPN)  
兴  
PigNet 44  
用  
位  
作为干网络  
( ROI) ,  
干网共享层  
( Region of interest alignROIAlign) ,  
支  
域  
线值计支并出  
ROI  
模  
Mask  
平均熵  
7 d 、  
的目失 连续  
28 d  
6
9. 6 kg  
左右仔猪取前  
集  
各不段  
2 500  
模式像  
型在达  
4∶ 1。  
PigNet  
作为集 训比例为  
明  
模  
本文形态 重的图  
Mask R-  
86. 15% ,  
上准达  
85. 40% 。  
本文与  
Mask R-CNN  
及其型进比  
CNN 11. 40 2. 12 sMask R-CNN  
30 ms。  
高  
分点 为  
短  
关键词 积神体  
: TP391 : A : 1000-1298( 2019) 04-0179-09  
:
;
;
;
;
;
中图分类号  
文献标识码  
文章编号  
Instance-level Segmentation Method for Group Pig Images  
Based on Deep Learning  
12  
1
12  
23  
4
12  
GAO Yun  
GUO Jiliang LI Xuan  
LEI Minggang  
LU Jun TONG Yu  
( 1. College of EngineeringHuazhong Agricultural UniversityWuhan 430070China  
2. The Cooperative Innovation Center for Sustainable Pig ProductionWuhan 430070China  
3. College of Animal Science and TechnologyCollege of Animal MedicineHuazhong Agricultural UniversityWuhan 430070China  
4. College of ScienceHuazhong Agricultural UniversityWuhan 430070China)  
Abstract: With the development of intelligence and automation technologypeople pay more attention to  
use it to monitor pig welfare and health in modern pig industry. Since the behaviors of group pigs present  
their healthy statusit is necessary to detect and monitor behaviors of group pigs. At presentmachine  
vision technology with advantages of low priceeasy installationnon-invasion and mature algorithm has  
been preferentially utilized to monitor pigsbehaviorssuch as drinkingeatingfarrowing behavior of  
sowand detect some of pigsphysiological indicessuch as lean yield rate. Feeding pigs at group level  
was used the most in intensive pig farms. Owing to normally happened huddled pigs showing in group-pig  
imagesit was challenging to utilize traditional machine vision technique to monitor the behaviors of group  
pigs through separating adhesive pig areas. Thus a new segmentation method was introduced based on  
deep convolution neural network to separate adhesive pig areas in group-pig images. A PigNet network  
was built to solute the problem of separating adhesive pig areas in group-pig images. Main part of the  
PigNet network was established on the structure of the Mask R-CNN network. The Mask R-CNN network  
was a deep convolution neural networkwhich had a backbone network with a branch of FCN from  
classification layer and regression layer to mask the region of interest. The PigNet network used 44  
convolutional layers of backbone network of Mask R-CNN network as its main network. After the main  
: 2018-10-17  
: 2019-01-29  
修回日期  
收稿日期  
:
基金项目 研发目  
( 2016YFD0500506)  
( 2662018JC0032662018JC0102662017JC028)  
和中创新金项目  
:
作者简介  
( 1974) , , , E-mail: angelclouder@ mail. hzau. edu. cn  
教授 博士 从事农业智能测与控研究  
180  
2 0 1 9  
networkthe output feature image was fed to the next four convolutional layers with different convolution  
kernelswhich was the resting part of the main network and produced binary mask for each pig area. As  
wellthe output feature image was fed into two branchesone was the region proposal networks ( RPN) ,  
the other was region of interest align ( ROIAlign) processing. The first branch outputted the regions of  
interestand then the second one aligned each pig area and produced class of the whole pig area and the  
background area and bounding boxes of each pig regions. A binary cross entropy loss function was utilized  
to calculate the loss of each mask to correct the class layer and the location of ROIs. Herethe ROIAlign  
was used to align the candidate region and convolution characteristics through the bilinear differenceand  
which would not lose information by quantizationmaking the segmentation more accurateand FCN of  
the mask branch used average binary cross entropy as the loss function to process each maskwhich  
avoided the competition among pig masks. After allthe ROI was labeled with different colors. Totally  
2000 images captured from previous five days of a 28-day experiments were taken as the training setand  
500 images from the next 6th to 7th day were validation set. The results showed that the accuracy of the  
PigNet on training set was 86. 15% and on validation set was 85. 40% . The accuracies on each data set  
were very closewhich showed that the model had effective generalization performance and high  
precision. The cooperation between the PigNetMask R-CNN ( ResNet101-FPN) and its improvement  
showed the PigNet surpassed the other two algorithms in accuracy. Meanwhilethe PigNet run faster than  
the Mask R-CNN. Howeverthe times of three algorithms spent on 500 samples of the validation set were  
similar. The algorithm can be used to separate individual pig from group-pig images with different  
behaviors and severe adhesion situation. The PigNet network model adopted the GPU operation mode,  
and used the three branches of classregression box and mask to compute parallel processing timewhich  
made the processing time of single image quickonly 2. 12 s. To a certain degreethe PigNet could  
reduce convolution parameters and simplify the network structure. The research provided a new  
segmentation method for adhesive group-pig imageswhich would increase the possibility of group-pig  
tracing and monitoring.  
Key words: group pig; image segmentation; instance segmentation; convolution neural network; deep  
learning; adherent pig body  
, ,  
明确情况取图中  
0
引言  
目  
视觉技术状况 评价利  
的 过 程 性 较 强  
21]  
、 、  
的重要技术之一 具有安装 平  
GIRSHICK  
R-CNN ( Regions with  
提 出 了 将  
( Region proposal) ,  
训  
,  
视觉技术较多猪  
CNN)  
域  
物和度  
1]  
测 主要用为 及猪  
2 - 3]  
4]  
22]  
度  
LONG  
量  
为  
猪  
5]  
6]  
CNN  
进 提 出 了 全 卷 络  
( Fully  
率  
情  
体  
, ,  
阶段 由于躺  
convolutional networksFCN)  
法  
, ,  
使用机视觉跟踪时 图存在  
要用来理解存在素  
。  
较多困难 要对个  
级别有研究员也将个  
23]  
。  
体进行追用 但群  
一  
面  
, ,  
中的不  
模式均占大  
7 - 8]  
重  
利于改  
中的体  
要用来分体  
, ,  
管理水平 的生产效率 从而提  
24]  
,  
此 解决连  
SILBERMAN  
2014  
年用来分析室  
全卷模  
( Fully convolutional instance-aware semantic  
25]  
LI  
问题视觉键  
景  
骤  
了分中的统  
segmentationFCIS) 。  
构  
, ,  
过卷积神提出感  
结合研  
9 - 13]  
( Convolutional neural  
( Region of interestROI) ROI  
积 神 络  
域  
个  
层处理提征  
14 - 20]  
networkCNN)  
ROI  
过对  
法  
背景知识晰  

全部评论(0)

暂无评论

上传资源 上传优质资源有赏金

  • 打赏
  • 30日榜单

推荐下载