Research On Cow Identity Recognition Based On Image Features And Deep Learning

With the improvement of living quality in China,the demand for dairy products has grown rapidly,becoming the industry with the fastest growing demand in the animal husbandry industry.The dairy farming industry has developed rapidly and the breeding scale has continued to expand.Management is becoming more and more important.Cow identification based on cow image features and deep learning is particularly important in modern animal husbandry,and can provide efficient and convenient modern management for the cow insurance industry and cow breeding management.In this paper,the image features of cows are integrated with deep learning.The latest deep learning model M2 Det model is used to train and recognize cow pictures,and to optimize and improve the problems in the experiment.The main contents of this article are as follows:(1)Production of cow data sets.Based on the cow pictures collected by a company in Jilin Province,a part of the cow pictures were randomly selected for screening and preprocessing to produce a cow dataset.Provide experimental materials for training of M2 Det model and identification of cows.(2)Experimental research on training of M2 Det model.The M2 Det model training in this article is trained with a homemade cow data set For the current cow recognition network model,the one-stage object detection network(such as DSSD,Retina Net,and Refine Det)and the twostage object detection network(such as Mask RCNN,Det Net)have certain limitations,so M2 Det is used in this article.Model(backbone network plus MLFPN)for cow image data training.Through experimental comparison,we can see that the detection performance is better than the existing technology.In the MS-COCO benchmark test,the VGG with the fully connected layer removed was used as the backbone network.When the single-scale reasoning strategy was adopted for M2 Det,41.0 APs were achieved at a speed of 11.8 FPS.When the multi-scale reasoning strategy was used,the AP was 44.2.The experimental results show that the detection effect of the M2 Det model is better than other network models.(3)Cow's identity recognition based on cow's image features and deep learning.a.Make a cow image dataset by labeling the collected cow pictures;b.Cow image data training.Set up the Image Net framework and experimental environment,input the cow image data set into the M2 Det model,train the cow image data set,and save the training data in log file format.c.Cow identification test.Only the pictures of the cows to be tested are imported into the network weights obtained by the training for forward propagation,the cows' data are compared,and the recognition results are output.The recognition rate of 800 cow pictures was 95.75%.The unrecognized pictures are mainly the cows' back flowers are not obvious,and the pictures are blurred.(4)M2Det model optimization and improvement.M2 Det also has a high accuracy rate in cow image recognition,but it has not achieved the expected results.Therefore,this paper enhanced the data set and made optimization adjustments to M2 Det.The pictures for cows are blurry and the back flower features are not obvious,which leads to unrecognized problems.This article first enhances the data set.Then the number of MLMPN TUMs and TUM channels were optimized,and the detection effect value of M2 Det was increased by 0.3.The last is to improve the backbone network of the M2 Det model.The HED network with richer low-level semantics is used as the backbone network,and then the improved experiment is performed.The training detection value AP obtained is increased by 1 from before,and the detection effect is improved.Experiment with the improved M2 Det model according to the cow's identity recognition process,the cow recognition was performed using 1,000 cow pictures,and the recognition rate reached 98.59%,which was 2.84% higher than the original M2 Det.