Research Of Face Recognition Method Based On Improved Low Rank Recovery Sparse Representation

The existing face recognition technology in the general case is simply, quickly and effici entl-y. But when affected by the uncertainty of shooting Angle and distance, the change of bright-ness, the variability of facial gestures,and the randomicity of noise etc,the rec ognition performance drops rapidly,thus restricting the development of face recognition field.In this paper, for the Sparse Representation based Classification(SRC) has certain defects, which is widely used in face recognition field, aims at finding better robust recognition method. In SRC: 1) use the poor performance of the unit matrix as the error dictionary in the progress of describing the noise and error the face image; 2)the dictionary incompletion caused by the insufficiency of the training samples; 3)Convex optimization problem in SRC is very complexity, thus not easy to spread. Aiming at those shortcomings of SRC, this paper presents a face recognition method Low-rank Recovery Non-negative Spare Representation-based(LRR_NSRC). The main idea of this algorithm is that using Low rank recovery(Low Rank Recovery, LRR) reconstruction matrix consisting of training samples to obtain a low-rank approximation matrix and a sparse of noise matrix that represents the error more effective and make them compose of a dictionary that Is more complete than the SRC, then add non-negative restrictions to the sparse representation to solve the high complexity of 1l norm optimization problem in SRC.Set up a different number of training samples on Yale B face database, respectively test SRC, ESRC, and LRR_NSRC. Results showed that when the training sample is less, the advantage is greater and the recognition rate is higher than SRC and ESRC.Then for LRR_NSRC is sensitive to changes of pose, while Transform InvariantLow-rank Textures(TILT) can directly extract certain invariant structures in 3D throughtheir 2D images by undoing the(affine or projective) domain transformations(That is tosay, TILT cast the quest for invariance directly as an inverse problem of recovering 3Dinformation from 2D image, thereby recover a low rank approximation matrix and asparse error matrix), proposes a improved algorithm TILT_NSRC. The improvedalgorithm transform(affine or projective) the training samples first, and then resumes corrected low-rank approximation matrix and error matrix, and at last, sparse representation and recognition.Set up a different inclination of test samples on FERET face database, respectively test SRC, LRR_SRC, and TILT_NSRC. Results showed that SRC and LRR_NSRC are more sensitive to changes in posture, while TILT_NSRC still robust on face images which have large variation on posture.