基于深度学习的羊脸检测与识别方法  

Method of Goat Face Detection and Recognition Based on Deep Learning

作　　者：魏斌 Masum Billah 王美丽[2] 尚诚 于建涛 姜雨[1] WEI Bin;BILLAH Masum;WANG Meili;SHANG Cheng;YU Jiantao;JIANG Yu(College of Animal Science and Technology,Northwest A&F University,Yangling,Shaanxi 712100,China;College of Information Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China)

机构地区：[1]西北农林科技大学动物科技学院,陕西杨凌712100 [2]西北农林科技大学信息工程学院,陕西杨凌712100

出　　处：《家畜生态学报》2022年第3期47-50,共4页Journal of Domestic Animal Ecology

基　　金：国家优秀青年自然基金(31822052);陕西省农业科技创新转化项目(NYKJ-2020-YL-07)。

摘　　要：家畜个体身份的精准识别对于动物行为研究和现代化选种选配具有重要价值,目前家畜身份的确定主要依据图像或视频中所存在的动物的脸部特征而展开。该文以羊脸识别为目标,提出了基于多种深度学习方法的羊脸检测与识别方案。对羊脸在图片中的位置进行框选和定位,将检测到的来自不同图片的羊脸区分开来,进行个体身份鉴定。采集来自同一群体的形态相近的35只奶山羊的共3121张图片,制作了羊的全身数据集和羊脸数据集。在羊脸框选定位阶段,当随机选取900张羊脸图片做训练,102张羊脸图片做测试时,使用深度学习YOLOv3算法检测到羊脸的正确率可达97%以上;基于可检测到羊脸数据集进行个体识别时,使用深度学习VGGFace模型仅能取得约64%左右识别准确率;而当选取其中616张正面羊脸作为训练集及198张正面羊脸作为测试集时,VGGFace预训练模型的识别准确率可达91%以上。试验结果表明,基于深度学习模型的羊脸检测效果较为理想,而羊脸识别工作则仅在正面羊脸上取得了较高的准确率,其它视角的羊脸识别工作仍有待进一步的研究。Accurate detection and recognition of individual identities of livestock is of great value for behavior research and breeding.Currently,the individual identity recognition of livestock is mainly based on the facial features of animals in images or videos.In this paper,goat face recognition was taken as the target,and a scheme of goat face detection and recognition based on multiple deep learning methods was proposed.It is divided into two steps:the first step is to select and locate the goat face in the pictures;the second step is to distinguish the goat faces detected from different pictures for individual identification.A total of 3121 pictures from 35 dairy goats with similar morphology were collected to make the whole body data set and the face data set.In the stage of the goat face detection,when 900 goat face pictures were randomly selected for training and 102 goat face pictures for testing,the accuracy of deep learning YOLOv3 algorithm could reach more than 97%.When performing individual recognition based on the detectable goat face data set,the use of the deep learning VGGFace model can only achieve a recognition accuracy of about 64%.When 616 frontal goat faces were selected as the training set and 198 frontal goat faces were selected as the testing set,the recognition accuracy of the VGGFace pre-training model reached more than 91%.The experiment shows that the goat face detection effect based on the deep learning model is accurate,while the goat face recognition only achieves high accuracy in the frontal goat faces.The goat face recognition from non-frontal view or non-full face pictures still needs to be further studied.

关 键 词：深度学习 迁移学习 羊脸检测 羊脸识别 

分 类 号：S826[农业科学—畜牧学]