Research Of Occluded Face Recognition Based On Sparse Representation

Face recognition has been an important technology in computer vision field,machine learning and biometrics.Although automatic face recognition has been widely studied in recent years, but due to the effect of occlusion, posture and lighting etc,the recognition performance still has a lot of restrictions.For a given facial image, use the database already face the test data confirm the identity of the person in the image, which is a basic feature recognition oper ations to be completed. Commonly used conventional face recognition methods are mostly poor anti-interference ability, and therefore subject to a certain degree of control in the application. In recent years, face recognition, sparse representation classification algorithm(Sparse Representation-based Classification, SRC) are more and more attention, which is a dictionary as a training sample, obtained in the test image sparse representation in the dictionary after factor solution for the minimal residual algorithm for identification. SRC face recognition method to provide a new direction, by a variety of experiments to prove that there is indeed a high performance on face recognition database, and for anti-interference factors are also relatively good performance. However, due to the presence of light and occlusion posture problems, recognition effect SRC becomes unsatisfactory; in particular, when the block area of more than 40 percent recognition effect significantly worse.For SRC in a challenging research topic-blocking recognition, this paper presents an improved algorithm based on block SRC dictionary, the training sample stage,and the normal face obscured face learning dictionary separately, the resulting occlusion dictionary each atom represents a block of type, so that you can use to block the dictionary test sample block represents a block sparse linear combination of various types, and two complementary interference dictionary, which only partially covered sparse block dictionary, sparse only that part of the non-blocking non-blocking in the dictionary.In addition paper also proposes improved algorithms SRC uneven block will face in accordance with a certain proportion of block sizes, each piece will find its minimum residual and minimum residual ratios according to SRC solving method, Finally, the results of each sub-block to give a summary of its longitudinal number,a unique value if the public is determined to be the face of it represents the kind of face, if not the only, then find a sub-block residuals than the maximu m one, it category as a result of the determination.Finally, the above two methods together to give shelter dictionary unevenness aliquot block SRC algorithm, in order to test the effect of its recognition, the three algorithms, respectively, for the same set of samples simulation experiments and found that no matter what type of real shelter, combining algorithm identifies the effects are more ideal, and in the higher dimension, to varying degrees better than the first two improved algorithms. In addition, to increase the difficulty of identifying the three algorithms, artificial occlusion random block experiments were carried out at the location and size are unknown beforehand, this paper the well algorithm has more satisfactory results, greatly improving the occlusion robustness.