Two-Directional Two-Dimensional Random Projection and Its Variations for Face and Palmprint Recognition

Abstract

Abstract. 2DRP (two-dimensional random projection) is two-dimensional extension of one-dimensional RP (random projection) to keep biometric images from being reshaped to vectors before RP for recognition. We propose a novel method called (2D) 2 RP (two-directional two-dimensional random projection) for feature extraction of biometrics. (2D) 2 RP directly projects the image matrix from high-dimensional space to low-dimensional space to extract optimal projective vectors at row-direction and column-direction. (2D) 2 RP, similar to RP, can also avoid the problems of singularity, SSS (small sample size) and overfitting; furthermore it has much less storage and computational cost than RP. Besides, the variations of (2D) 2 RP combined with 2DPCA and 2DLDA are developed. Experimental results and comparison discussion among (2D) 2 RP and its variations on face and palmprint databases confirm the performance and effectiveness of (2D) 2 RP and its variations.

Image recognition via two-dimensional random projection and nearest constrained subspace

“…This interest is somehow shared almost at the same time by A. Eftekhari et al [32] and L. Leng et al [33]. Although addressing the same problem, the focuses of A. Eftekhari, L. Leng et al [32,33] and ours are different.…”

“…Although addressing the same problem, the focuses of A. Eftekhari, L. Leng et al [32,33] and ours are different. Besides the theoretical analysis of 2D random projection and the assumption that 2D signal is sparse, A. Eftekhari et al reported a reconstruction algorithm of 2D sparse signal based on smoothed 0 -norm minimization.…”

“…For example, the 2D representation methods include 2DPCA [21] and its variants [22][23][24], 2DLDA [25], bilinear subspace learning [26,27], tensor analysis [28][29][30][31], and the recent common interest in 2D random projection [32][33][34].…”

“…Another category of 2D representation include those obtained by random linear projection [32][33][34]. Both categories are actually the order-two tensor analyses, which exploit the correlations among image pixels with different dimensions and in this sense is believe to lead to good classification performance for different applications such as image recognition and human gait recognition [28][29][30][31].…”

“…A reconstruction method of natural images (not explicitly sparse) and the problem of designing a cutting-edge classifier exploiting the 2D random projection features were not addressed in [32]. On the other hand, in [33], 2D random projection and its variations combined with PCA, LDA etc, were studied and compared with other feature extractors but without mention of the problems of 2D reconstruction and classifier design.…”