Some Models,Algorithms Of Representation Based Learning And Their Applications

Representational learning,as a hot topic of theoretical research in recent years,is widely used in many fields requiring feature recognition.Although the continuous innovations in theory and practice have solved some practical problems in life,there are still some challenges in the field of representational learning.1.In reality,many classification and recognition problems need to judge in a large number of categories.The number of categories affects the recognition performance of algorithms.2.Identification of many small size samples cannot fully utilize the characteristics and variations of all samples due to the small number of samples and the insufficient sample space.3.In practical application,it cannot meet the requirement of reasonable time while maintaining the recognition accuracy,because of the high time complexity.In view of the above shortcomings,this paper proposes three new representational classification methods.the specific work is as follows:1?To solve the problem of poor performance of representation based classification?RBC?methods on large-class-databases,a self-adaptive multi-phase linear reconstruction representation based classification?MPRBC?method is proposed.In each process,the reconstruction coefficients regularized by l₁-norm or l₂-norm are obtained at first.Then the similar classes are selected according to the sum of the representation coefficients in each class,and all samples of similar classes are retained as training samples for the next stage.This strategy finally produces a sparse class probability distribution with higher classification confidence.And the similar classes are selected adaptively according to the class coefficients,which improves the efficiency of the classification.2?Many representational learning methods are based on collaborative representation?CR?,but CR is represented on the dimension of samples,which cannot well reflect the aggregation feature of samples'classes.Probabilistic collaborative representation based classification?PCRC?,as a new extension of CRC,is classified from the dimension of samples'classes,so it has a natural aggregation property of classes.This paper proposes a multi-phase probabilistic collaborative representation based classification?MPCRC?method combining with PCRC and multi-phase method.This method is implemented in several stages,each of which uses PCRC to narrow the scope of training samples.In the last stage,more accurate classification results is obtained by fewer and similar classes.At the same time,a multi-phase weighted probability collaborative representation based classification?MWPCRC?method is proposed by introducing the idea of weighting samples.This method weighted the probability collaborative representation coefficients by the similarity of local distance between each query sample and all training samples.3?Many representational learning methods pay more attention to the similarity of the same class of samples,and in the subspace projection?SP?,they make more use of the differences of inter-classes,but both of them ignore the contributions of inter-classes.We can perfect the samples'space and reconstruct query samples more accurately by introducing contributions of inter-classes.In this paper,the non-negative representation and probabilistic collaborative representation are introduced to accurately describe the probability that all the training samples belong to each class,and then the classification of query samples is determined through collaborative representation,which can take full advantage of the characteristics and variations of all samples.The algorithm can be divided into two steps.First step is to obtain a probability distribution map of the training samples by non-negative representation or probabilistic collaborative representation.Second step is to determine the classification of query samples by collaborative representation coefficients which weighted by probability map.In addition,the training samples were expanded by adding symmetrical faces which can improve the recognition performance especially in small size samples.Experiments on different databases show that the proposed methods in this paper have better recognition performance and certain practical significance.