DAPM Model

The main step in fingerprint recognition is reliable extraction of pores. I have worked on two different pore extraction models. First of all a Dynamic Anisotropic Pore Model (DAPM) which adaptively determines the local ridge frequency and orientation required for pore model implementation. And a Difference of Gaussian model which using the difference between the two Gaussian convoluted images gives the extracted pores. 

The Dynamic Anisotropic Pore Model (DAPM) adaptively determines the local ridge features (i.e. ridge orientation and ridge frequency) necessary for the initial pore extraction stage. The following two sub-sections explain the overall DAPM pore modeling and extraction algorithm.  Open pores located on ridge and adjoining valley are anisotropic while closed pores are isotropic. The anisotropic pore model has a better pore detection efficiency irrespective of whether the pores are open or closed because along the orientation at pore, the intensity profile along the pore has a Gaussian shape.

Zhao et al. [1,2] proposed a dynamic anisotropic pore model which adaptively determines the scale and orientation parameter according to the local ridge features .i.e. ridge orientation and frequency. The dynamic anisotropic pore model is defined as follows:

Here Equation 1 is the reference model which models the pore located on horizontal ridges and Equation 2 is the rotated model which models the pore on ridges of orientation theta. The scale parameter sigma determined from ridge frequency is used for pore size selection and theta is used to direct the pore model along the orientation. Here the scale parameter is set as a constant multiple of local ridge period. Zhao et al. [3] proposed a pore matched filter for pore detection based on the automatic scale selection technique. 

Fig.1: Dynamic Anisotropic Pore Model

DAPM Pore Extraction Algorithm

The Fig.1 shows the flow of the DAPM pore extraction. The overall method with the example for reliably extracting pore is shown in Fig.1.2. The first stage is to compute the ridge orientation by dividing the image in a non-overlapping block (Fig.1.2 (b)). In second stage the mean ridge frequencies on all the blocks are estimated (Fig.1.2 (c)). It then proceeds to the third stage where fingerprint image enhancement is done via bank of Gabor filters obtained from ridge orientation and frequency. The enhanced image is binaries and its complement is obtained resulting in a binary ridge map where the ridge pixels have value 1. This ridge map (Fig.1.2 (d)) is used to remove spurious pores during post processing. In the pore detection stage the proposed pore model is applied to the image block-wise followed by applying a threshold to the image so that the pore pixels have value 1 and non pore pixel have value 0. This resultant binary image is the pore map. (Fig.1.2 (e)). The final pore map obtained is shown in the Fig.1.2 (f).

Fig.2:  Fingerprint pore extraction using DAPM (a) fingerprint fragment, (b) is    ridgeorientation of (a), (c) ridge frequency of fragment in (a), (d) Gabor filtered ridge map of (a), (e) initial pore map, (f) final detected pores.

                  

 Post-Processing

 As a post processing stage following steps were employed. First of all the pore map is filtered by the binary ridge map so as to remove the non ridge pixels. In this step, the pixels which are not on the ridges are removed. Then remove the pore pixels which have the least intensity. At last the connected components on the pore map are checked according to their size. The connected components having the number of pixels outside the pre-specified range (from 3 to 30 here) are removed from the pore map.

References : 

[1]   Qijun Zhao, David Zhang, Lei Zhang and Nan Luo, “Adaptive Fingerprint Pore Modeling and Extraction”, Pattern Recognition, vol. 43(8), pp. 2833-2844, 2010.

[2]   Qijun Zhao, David Zhang, Lei Zhang and Nan Luo, “Adaptive Pore Model for Fingerprint Pore Extraction”, IAPR 19^th International Conference on Pattern Recognition (ICPR2008), Tampa, Florida, USA, 2008. 

[3]   Qijun Zhao, Jianjiang Feng and Anil .K. Jain,” Latent Fingerprint Matching: Utility of Level 3 Features”, MSU, Tech. Rep, MSU-CSE-10-14, August 2010.