Kernel Sparse Representation-Based Classification Aided With Prior Information

Sparse representation has attracted considerable attention in many fields,such as,computer vision,signal processing,pattern recognition,image processing,and specifically in face recognition.The basic principle of sparse representation is to use as few atoms as possible to describe a signal in a super-complete dictionary,which provides a discriminative nature and has a relatively good reputation in both theoretical and practical scenarios.With the increasingly growing implementation of artificial intelligence in many real-life scenarios,the demand for faster and more reliable machine learning algorithms,especially for classification and target recognition,has been increasing.For this purpose,sparse representation-based classification(SRC)is one of the most recently proposed schemes,based on the emerging sparse representation theory.Thanks to the discriminative characteristic of sparse representation,the SRC demonstrates impressive classification performance relative to other well-known classification methods.Conventionally in SRC,a greedy approach such as orthogonal matching pursuit(OMP)is often used to obtain a sparse solution owing to its simplicity and also low time-complexity.In order to obtain a more accurate sparse representation,we exploit the prior information which represents favorable pattern in the data.This prior information is based on support probabilities.Such prior information indicates the probability of element in one certain position of the sparse vectors as nonzero.However,sometimes the prior information may be inaccurate,which may affect the performance.We can further improve the classification accuracy by addressing the issue of inaccurate prior information.In this work,we proposed logit weighted kernel orthogonal matching pursuit with joint judgment(LWKOMP-J)algorithm which is an extension of the existing logit weighted orthogonal matching pursuit with joint judgment(LWOMP-J).The LWKOMP-J algorithm aims to deal with incorrect prior information in high dimensional space,while LWOMP-J works in low dimensional space.In this algorithm,we used the RBF kernel function first to map the data from low dimensional space to high dimensional space so that classification becomes easier.Then in high dimensional space,prior information is added as an additive term while computing the correlation between the dictionary atoms and current residual.However,sometimes the prior information we use may also be inaccurate.In the LWKOMP-J algorithm,a specific judgment mechanism is used to achieve consistency between the support-set correlation and the prior information in order to reduce the negative effect of inaccurate prior information.Experimental results show that when compared with the existing LWOMP-J,the proposed LWKOMP-J algorithm enjoys better classification performance.Moreover,we extended the first scheme into a two-phase version to further improve the recognition rate and showing the efficacy of the proposed algorithm.The two-phase sparse representation-based classification(TPSRC)consists mainly of two phases of recognition,where the regularization weights become initiated in the first phase and updated during the next phase in order to increase the recognition rate.The reason for this increased recognition rate is because the TPSRC scheme not only emphasizes on minimizing the l2-norm between the test sample and the projection vector but also minimizing the l2-norm from this projection vector to the training samples.Experimental results also show that the proposed TPSRC performs better than the one-phase algorithm.