The Research And Application On Improvement Of The K Nearest Neighbor Algorithm About Image Classificaton

With the development of the computer and information technology,development and application of image recognition technology has been possible.One of the core issues to be solved in developing the identification system is to choose the classification method.The commonly used classification method includes genetic algorithms,neural networks,maximum entropy,SVM and k-nearest neighbor methods.Compared with classification method,k-nearest neighbor algorithm is more easy and it is more easy to implement.Therefore,the research on k-nearest neighbor methods has important theoretical and practical significance.On the basis of the summary of the k-nearest neighbor classification methods,this paper carries out a series of research method and its improvement centering on k-nearest neighbor method.Specific work is as follows.Firstly,in order to overcome this drawback,we propose a new approach,which is centroid neighbor classification algorithm based on local mean and class mean(CNNCM).CNNCM utilizes more effective centroid neighbor selection principle(NCN)to select k nearest neighbor points from the test sample points in each category of the training set.At the same time,it takes full advantage of the local mean and class means of k nearest neighbor points to determine the categories of test sample points.CNNCM not only have the characteristic of robustness which belongs to the non-parametric classification method based on local mean(LM),aiming at the problem of the outliers,and the effectiveness of neighbor selection in the limited samples for the principle of NCN,but also take full advantages of the impact of class mean on the classification.In order to verify the superiority of the proposed algorithm in the classification problems in which the accuracy of classification is ruled as evaluation criteria,the experiments for comparison between CNNCM and KNN as well as between KNCN and LM are taken in the 5 groups of real UCI datasets.The results show that the classification accuracy of CNNCM is obviously superior to other methods.Then aiming at the insufficient that the k-nearest neighbor classification method gives different neighbor samples the same contribution rate in the process of classification and is susceptible to be influenced by the outliers,the k-nearest neighbor algorithm of near centroid based on local weight(LWKNCN)is designed.LWKNCN takes advantage of the NCNprinciple to select k points near centroid from each category,and gives different weights to different centroid points according to the distance to the centroid point.A large amount of experiments is taken on 6 groups of real UCI datasets and synthetic datasets.The results show that the proposed algorithm in this paper is more efficient and reliable in the process of classification,the classification accuracy is obviously superior to KNN and KNCN algorithms.Finally,this paper uses CNNCM and LWKNCN to classify cotton and weed images.Specific work is as follows.Firstly,we use the method of weighted average and median filtering for image gray and filter out noise processing.Secondly,we segment images by adopting the method of between-cluster variance.Thirdly,5 kinds of operator of edge detection are used to edge detection of image.Forth,10 kinds of feature are extracted,which include the first moment,the second moment,and third moment of S component in the HIS color space model,energy,contrast and correlation selected from the textural features,and the length-width ratio,the circularity,the degree of rectangle,the properties of spherical shape selected from the shape features.We identify and classify the datasets of weeds by CNNCM and LWKNCN.The result shows that CNNCM and LWKNCN Has higher classification accuracy.