基于时空上下文信息的仔猪吮乳行为识别算法研究  

Research on piglet suckling behavior recognition algorithm based on spatio-temporal context information

作　　者：李颀[1] 马凯 LI Qi;MA Kai(College of Electronic Information and Artificial Intelligence,Shanxi University of Science and Technology,Xi'an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,西安710021

出　　处：《黑龙江畜牧兽医》2023年第24期51-56,147,148,共8页Heilongjiang Animal Science And veterinary Medicine

基　　金：陕西省农业科技创新工程项目(201806117YF05NC13(1));陕西省农业科技创新与攻关项目(2015NY028);陕西科技大学博士科研启动基金项目(BJ13-15)。

摘　　要：为了准确识别特征不明显且扎推聚群现象严重的仔猪吮乳行为,试验建立了一种基于时空上下文信息的仔猪吮乳行为识别算法,即先利用关键点检测技术定位侧卧母猪的哺乳区域,然后在哺乳区域中利用基于空间上下文信息模块的主干网络提取仔猪吮乳状态下的空间上下文信息,包括仔猪与母猪的相对位置和距离信息及仔猪嘴部与母猪乳房连接处的轮廓形状特征;为捕捉仔猪吮乳行为的运动特征,在主干网络的最后一层引入时间上下文信息模块,提取相邻帧间仔猪吮乳行为的时序特征;最后将特征输入到长短期记忆(long short term memory,LSTM)网络进行仔猪吮乳行为的预测与识别;以精确率、召回率和平均精度作为评价指标,将基于时空上下文信息的仔猪吮乳行为识别方法与Darknet53算法、原始YOLOv5主干网络+LSTM算法进行性能比较。结果表明:基于时空上下文信息的仔猪吮乳行为识别算法对仔猪哺乳行为识别的精确率、召回率和平均精准率分别96.9%、96.1%和96.3%,较Darknet53算法分别提高了14.7%、14.5%和14.4%,较原始YOLOv5主干网络+LSTM算法分别提高了12.5%、11.0%和11.3%。说明基于时空上下文信息的仔猪吮乳行为识别算法对仔猪吮乳行为有较好的识别效果。In order to accurately identify the suckling behavior of piglets with unclear features and severe clustering phenomenon,a piglet suckling behavior recognition algorithm based on spatiotemporal context information was established in the experiment.Firstly,key point detection technology was used to locate the lactation area of lateral sows,and then a backbone network based on spatial context information module was used to extract the spatial context information of piglets'suckling status in the lactation area,including the relative position and distance information between piglets and sows,as well as the contour shape features of the connection between piglets'mouths and sows'breasts.To capture the motion characteristics of piglet suckling behavior,a temporal context information module was introduced in the last layer of the backbone network to extract the temporal features of piglet suckling behavior between adjacent frames.Finally,the features were put into the long short term memory(LSTM)network for predicting and recognizing piglet suckling behavior.Using accuracy,recall,and average accuracy as evaluation indicators,the performance of the piglet suckling behavior recognition method based on spatiotemporal context information was compared with the Darknet53 algorithm and the original YOLOv5 backbone network+LSTM algorithm.The results showed that the accuracy,recall,and average accuracy of the piglet suckling behavior recognition algorithm based on spatiotemporal context information for piglet lactation behavior recognition were 96.9%,96.1%,and 96.3%,respectively,which were 14.7%,14.5%,and 14.4%higher than the Darknet53 algorithm.Compared with the original YOLOv5 backbone network and LSTM algorithm,the results were improved by 12.5%,11.0%,and 11.3%,respectively,indicating that the piglet suckling behavior recognition algorithm based on spatiotemporal context information had good recognition effects on piglet suckling behavior.

关 键 词：仔猪 吮乳行为 上下文信息 卷积神经网络 LSTM网络 

分 类 号：S828.9[农业科学—畜牧学] TP391.4[农业科学—畜牧兽医]