Cow Face Re-identification Based On Deep Learning

The large-scale,standardized,systematic,intelligent and accurate breeding and management methods are the most practical and forward-looking developing trends in the animal husbandry and aquaculture industries at present.In recent years,diary industry has gradually built a highly large-scale and standardized form.However,the intelligent and accurate breeding and management are still stay in primary stage.The reason is that the precise location and identification of cow as basic task are still at the research stage.The traditional identification methods,electronic devices methods and biometric methods cannot meet the needs of this basic task.Therefore,in this paper,taking cow face images data as research object,by using re-identification method,we construct a deep learning model to study the cow recognition task.The specific works are as bellows:(1)Based on a large amount of raw cow images data collected from different views and different environments by different image acquisition equipment(such as mobile phones,digital cameras),we contruct the cow face dataset through a series of operations liked primary cleaning,cow face extraction,image compression,image classification and data division.In the process of constructing,we has designed a cow face images extraction method combining with artificial way and network model,as well as a coarse cow face categories classification by network model and manual inspection after.The dataset contains 130,000 images totally,of which the train set contains 2,000 classes and 50 images of each category;the test set contains another 1,000 classes,the query set contains 10 images of each category and the gallery set contains 20 images of each category.(2)We train the fine-tuned model of current mainstream classification networks and state-of-the-art re-identification methods as well as perform re-identification experiments to them on the dataset we constructed in this paper.We compare and analyze the results of these experiments,find out the structual characters and advantages of these networks,preparing for subsequent network improvements.Among these models,the multi-granularity branch network MGN(Multiple GranularityNetwork)is the best recognition performance network,achieving 86.9% on Rank-1 accuracy and 89.1% on mAP performed on gallery with only one image of each category.(3)Based on the characters and advantages of networks used in the above experiments,we propose a global feature and local feature joint network named GPN(Global and PartNetwork).Our network takes ResNet as backbone network and own three branches,namely Middle branch that extracts global features of middle dimension from backbone,Global branch that extracts high-dimensional global features which pass through entire backbone,and Part branch that extracts unifid-size-block’s local features by using uniform division method.In addition the features extracted from the three branches are combined to use as the feature respresentation for re-identification comparison.The experiment results show that our network achieves 87.2% on the Rank-1 accuracy and 87.2% on the mAP is 89.1performed on the gallery with only one picture of each category.(4)We construct GPN-ST(Global and PartNetwork with Spatial Transform)by futher improving the Part branch of our GPN.The new network GPN-ST uses the STN module with attention mechanism to replace the original method of uniform divison,set up 4STN(Spatial TransformerNetwork)modules to extract different local features information of cow face.In this way it improve the re-identification performance of the network.Similarly,GPN-ST compares with GPN on the performances of the gallery with only one picture of each category,and the performance of GPN-ST in Rank-1 accuracy and mAP is2.8% and 2.2% higher than GPN,respectively.