Multi Local Means-Based Nearest Neighbor Pattern Classification

K-Nearest neighbor classification is a classical non-parametric statistical method in pattern recognition.Because of its simplicity,intuition and effectiveness,it is considered as one of the top ten data mining algorithms.However,the performance of nearest neighbor classification is sensitive to the choice of nearest neighbor parameter and the simple majority vote of classification decision.Its classification performance needs to be further improved.In recent years,local mean of nearest neighbor and representation of nearest neighbor have been applied to nearest neighbor classification,and the performance of nearest neighbor recognition has been improved.Aiming at the main problems of k-nearest neighbor classification,this paper studies some new methods of k-nearest neighbor classification based on local mean and representation of k-nearest neighbor.The main research contents are as follows:(1)In order to overcome the selective sensitivess of nearest neighbors,especially in the small sample size cases with noisy points,we propose a local mean representation-based k-nearest neighbor classifier(LMRKNN),which combines multi-local mean with representation-based distance.LMRKNN use multi local mean can make different neighbors have different classification contributions in different local mean.The multi local means of the nearest neighbors can not only overcome the selective sensitivess of the neighborhood size k in the small sample size cases,but also obtain more geometric and discriminant information.The distance based on multi local means not only enables multi local means to obtain adaptive classification weight,but also correctly reflects the similarity between test sample and training samples.The experimental results on real data sets and face data sets demonstrate the validity of the proposed LMRKNN.(2)With respect to the problems of the selective sensitivess of the neighborhood size k and the simple majority vote in KNN classification,we propose locality constrained representation-based k-nearest neighbor classification,including the weighted representation-based k-nearest neighbor rule(WRKNN)and the weighted local mean representation-based k-nearest neighbor rule(WLMRKNN).WRKNN considers the local information of each neighbor as a weight to constrain the representation coefficients of the corresponding neighbors.WLMRKNN uses local information of multi local means calculated by k-nearest neighbors as weight to constrain the representation coefficients of corresponding multi local means.WRKNN and WLMRKNN use representation-based distance to obtain adaptive weights for neighbors and multi local means,which can not only reflect the distribution of different local samples,but also further enhance the pattern discrimination.The experimental results on real data sets,time series sets and face data sets prove the validity of the proposed WRKNN and WLMRKNN.(3)According to the idea that the harmonic distance used as the classification decision in a new k-harmonic nearest neighbor classifider based on the multi-local means(MLMKHNN),a local mean representation-based k-harmonic nearest neighbor Classifier(LMRKHNN)is proposed.LMRKHNN firstly obtains the adaptive weights of each local mean based on the linear representation of multi local means based on nearest neighbors,and then uses the distance between test sample and the represented multi local means to design a new harmonic distance classification decision.On the one hand,multi local means can better overcome the selective sensitivess of the neighborhood size k.On the other hand,multi local means based on representation takes into account the contribution of the distribution of different neighbors to classification.The Experimental results on real data sets and image data sets demonstrate the effectiveness of the proposed LMRKHNN.(4)In order to further verify the effectiveness of the proposed classifiers in practical application,a prototype system of nearest neighbor image classification based on local mean representation is designed.In the prototype system,the proposed nearest neighbor classifiers are compared with other classifiers and validated on different image sets.The experimental results proves the practicability of the proposed classifiers.