沈明霞,太猛,CEDRIC Okinda,刘龙申,李嘉位,孙玉文.基于深层卷积神经网络的初生仔猪目标实时检测方法[J].农业机械学报,2019,50(8):270-279.
SHEN Mingxia,TAI Meng,CEDRIC Okinda,LIU Longshen,LI Jiawei,SUN Yuwen.Real-time Detection Method of Newborn Piglets Based on Deep Convolution Neural Network[J].Transactions of the Chinese Society for Agricultural Machinery,2019,50(8):270-279.
摘要点击次数: 1640
全文下载次数: 759
基于深层卷积神经网络的初生仔猪目标实时检测方法   [下载全文]
Real-time Detection Method of Newborn Piglets Based on Deep Convolution Neural Network   [Download Pdf][in English]
投稿时间:2018-12-21  
DOI:10.6041/j.issn.1000-1298.2019.08.030
中文关键词:  初生仔猪  实时检测  深层卷积神经网络  FPN算法
基金项目:政府间国际科技创新合作重点专项(2017YFE0114400)和国家自然科学基金青年基金项目(31802106)
作者单位
沈明霞 南京农业大学
江苏省智能化农业装备重点实验室 
太猛 南京农业大学
江苏省智能化农业装备重点实验室 
CEDRIC Okinda 南京农业大学
江苏省智能化农业装备重点实验室 
刘龙申 南京农业大学
江苏省智能化农业装备重点实验室 
李嘉位 南京农业大学
江苏省智能化农业装备重点实验室 
孙玉文 南京农业大学
江苏省智能化农业装备重点实验室 
中文摘要:针对初生仔猪目标较小、分娩栏内光线变化复杂、仔猪粘连和硬性遮挡现象较为严重等问题,提出一种基于深层卷积神经网络的初生仔猪目标识别方法。将分类和定位合并为一个任务,以整幅图像为兴趣域,利用特征金字塔网络(Feature pyramid network,FPN)算法定位识别仔猪目标;对比了不同通道数数据集以及不同迭代次数对模型效果的影响;该方法支持图像批量处理、视频与监控录像的实时检测和检测结果多样化储存。实验结果表明:在数据集总量相同时,同时包含夜间单通道和白天3通道的数据集,在迭代20000次时接近模型最优值。模型在验证集和测试集上的精确率分别为95.76%和93.84%,召回率分别为95.47%和94.88%,对分辨率为500像素×375像素的图像检测速度为53.19f/s,对清晰度为720P的视频检测速度为22f/s,可满足实时检测的要求,对全天候多干扰场景表现出良好的泛化能力。
SHEN Mingxia  TAI Meng  CEDRIC Okinda  LIU Longshen  LI Jiawei  SUN Yuwen
Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province,Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province,Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province,Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province,Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province and Nanjing Agricultural University;Key Laboratory of Intelligent Agricultural Equipment in Jiangsu Province
Key Words:newborn piglet  real time detection  deep convolution neural network  FPN algorithm
Abstract:Automatic recognition of newborn piglets has encountered several challenges such as small targets, ambient light variation, piglet adhesive behavior and object occlusion. A one stage DCNNs method was proposed to automatically and accurately recognize newborn piglets at high computation speed. The method merged classification and localization into one task and took the whole picture as the ROI of feature extraction, then using FPN algorithm to locate and identify piglets, which showed good generalization ability for natural multi interference scenes. The effects of different channel number data sets and different iterations on the effectiveness of the model were compared. Support for batch image processing, and real time detection of video and surveillance videos, with multiple storage of detection results. The recognition result of newborn piglets was output into three forms: video, picture and text file. The contents of the text included the number of piglets, the recognition confidence degree and the piglet coordinate. The combination of different output results could identify the state and behavior of piglets. The results showed that when the total amount of the data set was the same, the data set containing both night single channel and daytime three channel was close to the optimal value of the model at 20000 iterations. The precision of the model on the verification set and the test set were 95.76% and 93.84%, respectively, and the recall rates were 95.47% and 94.88%, respectively. The detection speed of the images with a resolution of 500 pixels×375 pixels was 53.19f/s. The video detection speed of 720P was 22f/s. The proposed system can meet the requirement of real time detection of piglets in a farrowing pen. 

Transactions of the Chinese Society for Agriculture Machinery (CSAM), in charged of China Association for Science and Technology (CAST), sponsored by CSAM and Chinese Academy of Agricultural Mechanization Science(CAAMS), started publication in 1957. It is the earliest interdisciplinary journal in Chinese which combines agricultural and engineering. It always closely grasps the development direction of agriculture engineering disciplines and the published papers represent the highest academic level of agriculture engineering in China. Currently, nearly 8,000 papers have been already published. There are around 3,000 papers contributed to the journal each year, but only around 600 of them will be accepted. Transactions of CSAM focuses on a wide range of agricultural machinery, irrigation, electronics, robotics, agro-products engineering, biological energy, agricultural structures and environment and more. Subjects in Transactions of the CSAM have been embodied by many internationally well-known index systems, such as: EI Compendex, CA, CSA, etc.

   下载PDF阅读器