Research On Single Sample Face Recognition Problem With Linear Regression And Fusion Of Gabor And LGBP

Face recognition is an important research direction in the field of computervision and pattern recognition. It has great difficulties due to the dimension of faceimage is too high, less training samples, and uncontrollable factors such as illumination,expression, partial occlusion. Especially when having only one training sample,training information is less and makes this problem even more intractable. Therefore,how to make full use of training information and priori information of training sampleis a critical problem.To these problems, this paper proposes a method of single sample facerecognition based on Gabor features of sub region, which divided face by thedistribution of human faces， extract Gabor feature for each sub region, thensub-regional recognition. The experimental results on the FERET, AR and ORL facedatabase demonstrated the effectiveness of this method.Later, Proposed the idea of fusion of feature based on sub-regional recognition,Further introduction of regional LGBP features based on Gabor feature of sub region,Experimental results on fusion of sub-regional features demonstrate that this methodget better performance than using just one feature.In addition, To take full advantage of prior information of single sample face, thispaper proposes the idea of sparse representation combined with linear regressionclassifier on problem of single sample face recognition. Firstly, Calculate the sparserepresentation of the training samples on the auxiliary dataset, selected severalneighbor samples in the auxiliary dataset, then calculate the changes within the class ofthese samples, using these changes combined with training samples to constitute themodels of training sample. Then these models are classified using linear regressionclassifier. Choose neighbor face with sparse representation, which not only considerthe geometry of the human face, but also conducive to identify the sample of samefacial expressions or gestures of training images. Further, within-class variation ofauxiliary samples is closer to the real change of training sample. Experiment resultsdemonstrated the effectiveness of this approach on AR and FERET face dataset.