Advanced Biometrics by David Zhang Guangming Lu & Lei Zhang

Author:David Zhang, Guangming Lu & Lei Zhang
Language: eng
Format: epub
Publisher: Springer International Publishing, Cham

8.3.4 Patch Based Reconstruction

In Sects. 8.3.2 and 8.3.3, we stretch the whole FKP image as a vector y for coding and reconstruction. In coding y over D by Eq. (8.6) or Eq. (8.8), we actually enforce that all the elements (i.e., all the pixels) in , denoted by y i ,  i = 1 , 2 ,  …  , n, share the same coding vector w over their corresponding sub dictionary (i.e., the ith row of D). The good side of such a global coding strategy is that the solution is very stable because only one global coding vector needs to be solved. The bad side of such a coding strategy, however, lies in its less flexibility because it does not allow the different parts of the FKP image to have different coding vectors. Considering the fact that different portions of a query FKP sample y may have different variations, it is reasonable to allow them to have different coding vectors so that the reconstruction can be spatially adaptive. Therefore, we can partition the FKP image into several patches, reconstruct separately each patch, and then combine them to obtain the whole reconstructed image.

If we partition the image into too many patches, the size of each patch will become small and the sub-dictionary corresponding to each patch will tend to be under-determined (i.e., the sub-dictionary will tend to be a fat matrix). This will reduce the stability of the coding process, no matter l 1 -regularization or l 2-regularization is used. Based on our experimental experience and considering the special pattern of FKP images, we partition the FKP image (size: 110 × 220) into 6 overlapped patches, as illustrated in Fig. 8.5. Four patches of size 60 × 45 lie in the four corners of the FKP images, and two patches of size 60 × 155 lie in the middle of the image. The reason that we set two fat rectangle patches in the middle is based on the observation that the pose variation along vertical direction has bigger effect than that along horizontal direction on the FKP recognition accuracy. Therefore, pay more attention to the pose variation along vertical direction in the partition.

Fig. 8.5The patch partition of FKP images

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.