Research On The Design Algorithm Of Classifier With High Correct Recognition Rate Based On Package Set

Great achievements have been made in the design of classifiers over the past decades.However,some classifiers with high correct recognition rate,such as SVM,SVDD and deep learning classifiers,still have about 2% error recognition rate.These classifiers cannot be directly used in high-precision authentication and recognition situations as serious diseases,self-identities,banknotes,bills and terrorists authentication and identification,because the error recognition rate of them needs to be close to 0%.Aiming at the above problems,this paper proposes and researches on a design algorithm of the high accuracy classifier with the appropriate rejection mechanism,based on the method of superball support vector machine and the cover concept of bionic pattern recognition theory.The algorithm mainly includes:(1)the constructive algorithm of the intensive package set in the similar characteristic set;(2)the solving algorithm of the intensive package surface in the similar characteristic region based on the similar characteristic set and the intensive package set.(3)The method of setting all public areas outside the tight wrapping surface as rejection areas of the classifier.The above algorithm solves the problem that the similar characteristic set determined by traditional classifiers will seize the known or unknown characteristic set,and introduces a rejection mechanism,which can give a more accurate description of the similar characteristic set under the condition of low rejection rate,thus significantly reducing the error recognition rate of the classifier.In this paper,the problem of handwritten digits recognition is taken as an experimental verification.Under the software environment of MATLAB R2016,the expanded MNIST dataset table is used to carry out the experimental test.Experiments show that under the premise of low rejection rate,the error recognition rate of the classifier designed in this paper is close to 0%,and the correct recognition rate is close to 100%.That is to say,the classifier designed in this paper can be applied to high-precision authentication and recognition situations.