Fast Face Recognition Via Collaborative Representation

Face Recognition(FR)has always been a hot topic in machine learning and lots of algorithms have been proposed for different scenes and targets.As Compressed Sensing Theory(CS)has grown and matured,one application in FR is Sparse Representation Classification(SRC),which is insensitive to feature extraction and robust to occlusion.In this paper,I concentrate on an algorithm called Collaborative Representation Classification(CRC),which is based on SRC.CRC inherits the robustness to occlusion and dramatically improved the speed of recognition.In order to improve the accuracy and practicality of CRC,I come up with three schemes:(1)Use multi-features as input.The algorithm uses different features to train different models and then calculates the sum of weighted residuals under the same label.This improvement can exploit different advantages of each features and bring the increase of accuracy.(2)Use Weighted Relative Distance(WRD)to judge outliers.The sparsity of CRC is weaker and Sparsity Concentration Index(SCI)is no longer useful to judge outliers.Independent of the code's sparsity,WRD considers both distance and similarity of optimal and second-best solutions,which performs better than SCI in CRC.(3)Use transform dictionary to solve the problem of lack of samples.Both SRC and CRC perform badly when samples is insufficient.To solve this problem,the algorithm extracts basis vectors under different illuminations/expressions/occlusions from standard face gallery and then forms a transform dictionary(TD).With the help of TD,the incomplete dictionary generated from training set can represents all face images in different scenarios with only a few samples.