Title:

Supplementary Materials

Verification of the MMFRTN feature

The proposed MMFRTN feature is also tested on the Hong Kong Polytechnic University Palmprint DatabaseⅡ(PolyUⅡ). PolyUⅡdatabase contains 7752 grayscale palmprint images from 386 palms corresponding to 193 individuals. In this database, about 20 samples from each palm were collected in two sessions, where about 10 samples were captured in the first session and the remaining 10 samples were captured in the second session. In our experiment, the first 3 samples captured in the first session are chosen as training samples, and the 10 samples captured in the second session are the testing samples. For comparison, we use some other block-wise histogram features, such as the Local Ternary Pattern (LTP), MFRT based LTP, and PCANet [29].

The global feature fglobal is used as output of MMFRTN. In MFRT filtering stages, we selected filter banks with four different sizes of 15x15, 19x19, 23x23, and 27x27. For each filter size, the filter bank is composed of filters along 12 different directions of 0°, 15°, 30°, 45°, 60°, 75°, 90°, 105°, 120°, 135°, 150°, and 165°. In each MFRT filtering stage, magnitude map and phase map can be used separately or combined together.

For the LTP, we use the uniform encoding LTPu2 as LBPu2. For MFRT based LTP, we implement different MFRTs before LTP and use the best experimental result. For PCANet, we select a set of parameters according to [29] and also use the best result. In the block-wise histogram generating stage of these methods, we segment the entire image into 8x8 non-overlapping blocks. In addition, the linear SVM is used as the classifier for all these methods. The error rates on PolyUⅡdatabase are listed in Table S-1.

From Table S-1 we can get the following findings. First, by combining with the MFRT, MFRT based LTP can achieve much better result than LTP. Second, using the magnitude map and phase map separately can get lower error rates than PCANet. Finally, the error rate can be further reduced by combination of the magnitude map feature and phase map feature. The experimental results on palmprint database demonstrate that the proposed MMFRTN can effectively extract the line-like feature.

Table S-1. Comparison of error rates (%) on PolyUⅡdatabase

Download Area:

1. Image datasets:

Hong Kong Polytechnic University Palmprint DatabaseⅡ(PolyUⅡ)： http://www4.comp.polyu.edu.hk/~biometrics/

2. Source codes: (matlab code, coming soon...)

Return