Low-Resolution Face Recognition Based On (2D)~2PCA

Eigenface-Domain super-resolution for face recognition is an effective method to recognize the low-resolution faces. This method transfer the super-resolution reconstruction from pixel domain to a lower dimensional face space (the Eigenface-Domain). Such an approach has the advantage of a significant decrease in the computational complexity of the super-resolution reconstruction and can get a ideal reconstruction results.In this method, dimensionality reduction technique in face recognition is the principal component analysis (PCA). As we know, PCA need to align images into 1D vector before extract bases. It maybe induces high computational complexity of PCA as it is related to the size of image vectorization. Moreover, matrix-to-vector transform may cause the loss of the intrinsic spatial structure of the original image. For covering the drawbacks, researchers established the new methods based on PCA.In this paper, two new super-resolution approaches for face recognition is proposed. The first one is proposed based on the method of two-directional two-dimensional PCA ((2D)~2PCA). The main advantages of improved approach is that can obtain the better reconstruction and recognition rate. But the problem is that the result of (2D)~2PCA is a matrix but the original super-resolution observation model apply to vector. So, super-resolution reconstruction need to align images into 1D vector.For covering the problem, the second new approach is proposed based on the 2D observation model, which can recognize a matrix. The experimental results on AR database, ORL database, and YALE database revealed that, the modified algorithms have better reconstruction results and recognition rate than existing algorithms.