| The Information Society is developing rapidly.As an important basis for the information management of cows,cow individual recognition has become a hot research focus in the development of animal husbandry.With the deepening research in the field of artificial intelligence and pattern recognition,it has gradually become an important measure to promote the cow information management that using image processing algorithms to link cow individual recognition with computer vision.However,the key technology of image recognition lies in the design of feature extraction and classification algorithms.In order to effectively complete the feature extraction and classification for the images of cow individuals with black and white pattern,and improve the accuracy of cow individual recognition,the main work in this paper follows:Firstly,we describe several feature extraction algorithms used widely in the field of the cow individuals’ images recognition,and give the corresponding feature maps and the analysis of advantages and disadvantages about those algorithms.Besides,we put forward an improved Kernel Principal Component Analysis algorithm.After that,we introduce the Support Vector Machine(SVM)classification algorithm and related theories of the kernel functions,and give the schemes of solving multi-class problems.Secondly,aiming at the limitations of traditional algorithms for images recognition we point out the advantages of deep convolutional neural network in extracting image features.By combining the characteristics of cows’ black and white pattern information,in this paper we propose a convolutional neural network model named as CowNetFull.Compared with the existing individual identification cow models,our model adds layers and replaces the mode of fixing the size of convolutional kernels with the way of flexibly setting values for convolutional kernels according to different layers.In addition,Relu is chosen as the activation function of the model in this paper to accelerate the convergence process of the network,and at the same time,the Dropout layer is introduced behind each fully connected layer,to avoid the phenomenon of over fitting because of too many parameters.Thirdly,on the basis of the CowNetFull model,we propose an algorithm for cow individual recognition named as Haar-CowNet-SVM.By analyzing the shortcomings of the traditional classification of convolutional neural network,the idea of combining the CowNetFull’s ability of feature extraction with SVM’s ability of classification is brought up.In addition,considering the performance of our network,we add Haar wavelet transform as pretreating operation,and integrate the sub images with reasonable weights to preserve the effective information of the image,to eliminate some simple noises and to reduce the dimensionality.Moreover,in order to solve the problem of insufficient data for cows,we build up a small standard database for cow individual recognition experiments during from collecting data to processing data.It is a database of 30000 images of 30 cows and 1000 images for per head,of which 20 cows’ images constitute the training set and validation set,the remaining 10 cows form the test set and each cow individual contains pictures of different angles,to ensure the diversity of samples.Finally,some experiments and results are given in the experiment part,such as the setting of the parameters,the selection of the classification algorithm,the contrast of the different recognition models and so on.After the Haar wavelet preprocessing operation,CowNetFull model is trained by 20000 images of 20 cows and then the cow individual recognition accuracy of the validation set reaches up to 98.8% with SVM,significantly higher than the other kinds of algorithms.Besides,in order to verify the strong generalization ability of the proposed feature extraction algorithm in this paper,we draw the map including Receiver Operating Characteristic Curves(ROC)for different algorithms according to the experiments results on the test set. |
PDF Full Text Request