Research On Sparse Representation Classification Algorithm And Its Application

The effectiveness of many face recognition algorithms relies heavily on large-scale and representative training sample set. However, in fact, single sample face recognition problem is more common. By using the sample extension method to make a single sample problem into sample enough problem is of particularly importance. Samples get by the existing sample expansion methods are highly correlated to the original data which makes the recognition rate of single sample problem not good enough yet. The differences between different people are to some extent similar, so differences caused by light, expression and shade can be shared among different people. The intra-class differences can get from the sample enough database to enlarge the single sample training data. The ESRC algorithm extracted intra-class differences to supplement the single sample data from database with sufficient samples by subtraction of images in the same class. At last the SRC was used to classify test samples and the recognition rate on AR database was increased by20%to30%. The B-JSM could extract the common features and differences of images from the same class. The difference images were extracted by B-JSM as the extra images to enlarge the original single sample data. At last, the SRC was used to classify the test samples. Experiments were conducted on AR database and the recognition rate was improved about35%which proved the B-JSM is an more effective sample expansion method than ESRC and provided a new way in solving the single sample problem.Although, sparse representation theory is widely used in image compression, image denoising, feature extraction, image retrieval, image restoration and many other fields, but it’s reconstruction algorithm is too complex and it has high demand on signal sparsity. CR based classification with regularized least square algorithm combined collaborative representation with sparse limitation, which were the two most important points in SRC. The algorithm speeded up SRC and also got comparable results to SRC. In this paper a new contribution function based classification algorithm was proposed based on CRC RLS. The contribution function counted the contribution value of each class when images in this class encoded the test sample. The bigger the value was, the more likely the test sample would belong to the class. The algorithm was applied to person-independent facial expression recognition and obtained recognition rates72.38%in JFFE database and79.43%in CK database. The experiment shows the proposed contribution function based classification algorithm can widened distance between similar expressions and make the expression recognition more accurate.