Research Of Sparse Represention Face Recognition Algorithm Based Image Feature

Collecting face images is often influenced by the changes of illumination, facialexpressions and postures; therefore, it is difficult to meet people’s expectations inpractical application. Sparse image-based face recognition is a new method to put thecompressed sensing theory into pattern recognition, and it possesses strong robustnessin overcoming illumination, facial expression, and posture changes. With regard to theabove issues, the thesis conducts in-depth study on sparse representation-based facerecognition on the basis of related research achievements home and abroad and thelatest research progress.Firstly, applying the principal component analysis into the experiments on thenearest neighbor and K-nearest neighbor classifiers, the paper conducted a detailedanalysis of the algorithm by adding rejected situation from the practical application. Inthis way, the experiment obtains the value of the rejected parameter as well as otherperformance index in rejected situation.Secondly, the sparse representation for face recognition generally applies thefacial features by undersampling for this kind of face recognition. This dissertationmakes some improvements through using the well-descriptive characteristics of PCAon face images to conduct feature extraction of face recognition and apply the overallcharacteristics to sparse representation of face recognition. The results achieved bettereffect compared with undersampling, which shows that the robustness ofEigenface-based sparse recognition is stronger than that of the undersamplingcharacteristics.Finally, based on the fact that the Gabor wavelet can utilize local features insparse representation by extracting the fine facial features from multi-scales andmulti-directions, this thesis puts forward partitioning sub-band of Gabor wavelet basedsparse representation for face recognition. This algorithm let the Gabor wavelettransform at different scales and orientations, and carry out blocks for every sub-band obtained. Then, blend energy information of every sub-band into its eigenvector, andcombine each sub-band eigenvector together to get intense Gabor eigenvector. At last,such eigenvector is applied to sparse representation for face recognition algorithm.The result of the experiment indicates that the recognition rate of this algorithm ishigher than that of using of the Gabor transformation, and this algorithm possessesstronger robust when dealing with illumination, facial expressions and postures.Considering the practicality, the algorithm is added the rejected analysis, and theexperiment shows that in the presence of the rejected performance, the proposedalgorithm is still better than the sparse representation of the face recognition algorithmsimply based on the Gabor transformation.