Research Of Individual Dairy Cattle Recognition Based On Feature Fusion

The rise of internet has led to the gradual industrialization and intelligent development in various industries.With the deepening of researches in artificial intelligence and pattern recognition,it has gradually become an important measure to promote the informationized management of large-scale dairy cattle in order to link individual identification with computer vision by using image processing algorithms.And the key of image recognition and classification is to choose and design image feature extractors and classifiers.Previously,the individual identification of dairy cattle mainly depends on artificial observation mode.But this mode has high rate of false recognition.After that,it has developed into the mode of label recognition,but the labels are easily damaged and cause too much daily interference to dairy cattle,which may reduce the milk production.In recent years,automatic identification of individual dairy cattle has become the universal mode,which can extract the image characteristics of dairy cattle automatically,accurately and efficiently.In order to effectively complete the feature extraction and classification for the images of cow individuals,and improve the accuracy of cow individual recognition,the main work in this paper is as follow:Firstly,LBP algorithm,HOG algorithm,Harris algorithm and SIFT algorithm are adopted to extract the features of head parts of cows.PCA,LDA and LPP are used for dimensionality reduction respectively.The cows' classification experiments are carried out with SVM,KNN and random forest algorithm,and the shortcomings of each of these algorithms are summarized.Secondly,according to the comparison results of the algorithms of features extract,in this paper an improved LBP algorithm and an optimized HOG algorithm are proposed to form a fused feature vector.The improved LBP algorithm updates the LBP code by comparing adjacent pixel values,and uses the equivalent pattern to reduce the feature vector dimension.Then the improved LBP algorithm uses different blocks to calculate the histogram of the image distribution.The optimized HOG algorithm uses a simple template to calculate the direction,gradient and size of the gray scale,which also removes the trilinear-difference step and reduces the time of the feature extraction.Thirdly,we analyze the principal component of the extracted feature vectors which have been fused.Next the dimensions of these extracted feature vectors are reduced.By comparing the experiment results of several classifiers,based on the SVM classifier which has shorter classification time and higher classification recognition rate,the network-search algorithm is used to optimize the SVM classifier in order to find the best parameters for classification and recognition.In addition,for the purpose of shortening the extraction time and enriching the extraction information of cow-head-part image features,normalization and high-passfiltering pre-processing operations are adopted before the feature extraction.Finally,20 dairy cattle are taken as sample,with the training set containing 16000 eigenvectors and the testament set containing 4000 eigenvectors.The correct rate of recognition can reach 99%,which verifies the effectiveness and feasibility of the proposed algorithm.Meanwhile in the experimental part the discussion of the parameters involved in the algorithm ensures the rigor of the results.