Classification via two layers sparse representation

Recently, sparse representation based recognition (SRC) has been widely used and made great success in face recognition. SRC first represents a testing face image by a sparse linear combination of all the training images, and then classifies the testing sample by evaluating which class leads to the minimum representation error. However, just choosing the minimum error as the rule of classification is usually not robust for noise, gesture varieties and illumination as the images built by the true class may be disturbed and the error may be bigger than the false class. What´s more, sparse coding is a collaborate representations process, so it tends to get the wrong way when coding the test images. This paper introduces a two layers classifier to get the labels: the first layer chooses the labels of the n minimum errors rebuilt by SRC, and the second uses some classifiers (e.g., NN or NS) to get the true label in the n classes. Through our experiments on face 94 and AR database, the recognition rate is improved by five or more percent.