Design And Implementation Of Classification Algorithm Based On Discriminative Low-Rank Representation

Face recognition is a hot issue in pattern recognition field.Classification algorithm design is a basic problem in face recognition research,and feature representation is the key to affect the performance of classification algorithm.Traditional feature representation methods are often susceptible to noise such as masks,illumination,expressions,and gestures in face images.Recently,researchers have proposed low-rank decomposition and low-rank representation methods to eliminate noise interference in face images,so as to obtain effective discriminant feature representation.Aiming at the problem that the performance of traditional face classification algorithms is susceptible to noise such as mask,illumination,expression and posture,this paper applies low-rank representation to traditional classification algorithms and designs and implements several face classification algorithms with strong discrimination.The main work of this paper is as follows:1.Two-phase low-rank representation classification methodIn order to improve the classification performance of existing low-rank representation classification methods,we propose a two-phase low-rank representation classification method(TPLRRC).The algorithm is divided into two phases.The first phase is to use the low-rank representation classification algorithm to calculate the Mclasses most similar to each test sample;in the second stage,a new training sample set is constructed using the above Mclasses,and then a low-rank representation classification algorithm is applied again on the new training set to solve the category of each test sample.Experiments verify the effectiveness of TPLRRC algorithm on AR and Yale B face database.2.Weighted low-rank cooperative representation classification methodThe low-rank cooperative representation algorithm obtains a low-rank representation by a kernel norm minimization constraint method,and treats each singular value equally in the constraint.In practical applications,different singular values have different contributions to the final feature representation,but existing low-rank collaborative representation methods do not consider their different contributions.In response to the above problem,we propose a weighted low-rank cooperative representation classification method(WLCRC).The algorithm uses the weighted kernel norm instead of the ordinary kernel norm to ensure that different singular values can be adaptively assigned different weights,so that the finally obtained feature representation is more discriminative.Experiments on AR and CMD PIE databases verify the effectiveness of the WLRRC algorithm.3.Non-negative sparse low-rank representation classification method with inconsistent structureAiming at the problem that the existing low-rank representation classification method has unsolvability and low discriminative power,we propose a non-negative sparse low-rank representation classification method(NSDLRRC).The algorithm adds non-negative,sparse and structural inconsistency constraints on the basis of low-rank constraints,so that the finally obtained feature representation has better interpretability and stronger discriminating ability.Experimental results on AR and Yale B face databases show that the NSDLRRC algorithm is superior to the existing low-rank representation classification method in terms of interpretability and classification accuracy.4.Face Detection and Recognition SystemWe designed a face detection and classification recognition system.The system is divided into three parts.The first part is to preprocess the face image,including binarization,affine transformation,histogram equalization,etc.The second part is face detection,using AdaBoost algorithm to detect face images.The third part is face classification and recognition,using the algorithm in this paper for classification and recognition.