Research On Behavior Recognition Of Dairy Goat Based On ResNet Network

Behavior identification and analysis of dairy goat can be used as direct evidence of the welfare and health status of dairy goat.Behavior identification provides guarantee for the intelligent and precise breeding of dairy goat.This article takes dairy goat video as the research object,and uses the deep residual network model ResNet to realize the recognition of dairy goat standing,lying down,curling up and eating behavior.The main research contents and conclusions of this article are as follows:(1)Acquisition and production of dairy goat data set.In order to provide a sufficient number and variety of dairy goat data set,first of all,a remote video monitoring device is set up at the dairy goat farm to shoot the dairy goat video,the video is preprocessed,and the initial dairy goat image data set is produced.Secondly,three data enhancement schemes of geometric transformation,geometric and color transformation and multi-scale transformation are used for data enhancement,and finally three dairy goat data sets are produced.Through comparative analysis,data enhancement based on multi-scale transformation not only effectively prevents the occurrence of overfitting,but also maximizes the accuracy of model recognition.(2)Constructing a dairy goat behavior recognition model based on ResNet.Using the convolutional neural network to independently learn image features and perform classification and recognition features,based on the deep residual network ResNet,18-layer,34-layer and 50-layer dairy goat behavior recognition models are constructed.Through experiments,not only the ResNet network model is verified in the field of dairy goat behavior recognition,the feasibility of the 50-layer deep residual network ResNet-50 has been verified as the model with the best recognition effect,and the recognition accuracy rate can reach 91.1%.(3)Constructing a dairy goat behavior recognition model based on the improved ResNet-50.In order to further optimize the dairy goat behavior recognition model and achieve a higher recognition accuracy,this paper makes improvements to the ResNet-50 network model.Firstly,compare the adaptive learning rate algorithms Ada Grad,RMSProp and Adam,improve the optimal algorithm Adam to the Adam New algorithm,introduce Nesterov momentum and improve its modified second-order moment deviation formula to control the dynamic learning rate of the algorithm,and ensure its monotonically decreasing adaptation to avoid the problem of large and small learning rate,the Adam New algorithm is used to optimize the model,experiments show that the improved algorithm has better convergence;secondly,the residual structure of ResNet-V1 and ResNet-V2 is compared and the ResNet-V2 structure is used to optimizing the model,the accuracy of the model is improved to about 93%;Thirdly,the activation function Re LU and SELU are compared and compared.The activation function SELU is used to replace Re LU and the model is optimized with the lecun＿normal initialization weight method.Finally,the improved ResNet-50 dairy goat behavior recognition model recognition accuracy rate increased from 91.1% to about 95%,an increase of 3.9%.In summary,this study based on the improved ResNet-50 dairy goat behavior recognition model realized the recognition of dairy goat’s standing,lying down,curling up,and eating behavior with high accuracy,providing guarantee for the healthy breeding of dairy goats.