Cattle Identification In Complex Scenes

With the promotion of modern animal husbandry in China,it is particularly important to establish smart pastures to achieve fine management through Internet,big data,5G communication and other technologies.It is a key step to achieve smart pastures through cattle identification.The wide application of biometric technology based on deep learning has led to the rapid development of smart ranching combined with deep learning.However,in the application of practical smart ranch,due to the diversity of ranch environment,the collected cattle body images have a large difference in posture,which leads to difficulties in multi pose recognition and imbalance of data sets.In this thesis,cattle images are mainly collected through the cowshed and milking channel.In the cowshed,cattle have a wide range of activities,and the collected images of each cow show diversity of postures.It is difficult to identify cattle in different postures;In the milking channel,the cattle are limited in the narrow channel,and the cattle pass through quickly.The collected image posture of each cattle is relatively simple,and the collected effective samples are few.When the cattle images in the cowshed and milking channel are mixed for training,the sample imbalance phenomenon occurs.This thesis conducts cattle recognition based on the body pattern on the back of cattle,namely cattle body recognition.The main research contents are as follows:1.Since there is no public cattle body image data set,the data set made by our laboratory is used.In this thesis,the data set of 389 cattle was collected from 2 cattle farms,mainly through the cattle shed and milking channel.In the cowshed,309 kinds of cattle body data were collected by the camera deployed above the cowshed.The cattle in the cowshed can only move freely,and the other is tied to the cowshed with a rope.The image attitude of the cattle in the cowshed is quite different.In the milking channel,the camera deployed above the milking channel is also used to collect data,and a total of 80 cows are collected.The range of cattle in the milking channel is small,so there is little difference in the attitude change of cattle images.2.Aiming at the difficulty of cattle multi pose image recognition,Transformer cattle body recognition algorithm based on local feature aggregation is proposed.First of all,fully considering that the overall difference of cattle body images under different postures is large and the correlation of local feature information of cattle body exists,therefore,the correlation of different parts of cattle body is established by using Transformer’s multi head self attention mechanism.Secondly,the convolution space aggregation module is used to increase the local feature information that is not available in Transformer to improve the representation ability of feature information.Finally,triple loss is used to expand the feature distance between cattle with high similarity,and label smoothing cross entropy loss is used to reduce over fitting and establish a certain correlation between cattle.The experimental results show that the algorithm effectively improves the recognition performance of multi pose cattle images.3.Aiming at the imbalance problem of cattle body data set,a bilateral identification algorithm based on MBN Transformer for unbalanced cattle body data is proposed.First,the images of the balanced sampler and the random sampler are mixed and enhanced by the image mixing enhancement module,and the image mixing degree is adjusted by dynamically fusing the mixing parameters to reduce the over fitting phenomenon of cattle in the milking channel.Secondly,the Transformer encoder is used to design the conventional branch and the balanced branch,respectively processing the image mixed enhancement data of the cowshed random sampler and the image mixed enhancement data of the milking channel balanced sampler,and the output features of the two branches are fused through the dynamic balance factor,so that the final fused output features change with the training times to solve the problem of poor recognition performance of cattle data with single posture in the milking channel.The experimental results show that the algorithm significantly improves the recognition performance for unbalanced cattle body data.4.Aiming at the deployment problem of the actual scene of cattle body recognition algorithm,the research of cattle body recognition algorithm platform based on AI box is proposed.According to the actual ranch environment,the overall architecture of the entire algorithm platform is designed,including camera deployment,algorithm model shift,feature comparison and template update.