Experimental Results of DoG

So far we have seen the experimental results of DAPM and here we will be seeing the experimental results of DoG pore model. Also the experiment was conducted on a high resolution partial fingerprint database using the proposed pore extraction and pore matching methods and implemented in the MATLAB software's image processing toolbox.

  

Pore Detection Accuracy

We first assess the pore detection accuracy of the Difference of Gaussian. For this purpose we choose a set of 10 fingerprint images from the database. Fig.1 shows the examples of pore extraction result of the two pore extraction model. In addition to the visual evaluation of the pore detection results, we calculated the pore detection accuracy on the 10 fingerprint images by using the two metric: True Detection Rate (TDR) and False Detection Rate (FDR).

A good pore extraction should have a high TDR and a low FDR simultaneously. Table 1 list the average detection accuracy of the two pore extraction considered as a true pore whereas the one which lies outside the ridge i.e. on valley are considered as false pores.

Table 1: Performance Measure of Difference of Gaussian

Images	FDR	Percentage	TDR	Percentage
Image_1	42/473	08.87 %	276/473	73.15 %
Image_2	41/389	10.53 %	338/389	86.88 %
Image_3	38/448	08.48 %	359/448	80.13 %
Image_4	29/422	06.87 %	341/422	80.80 %
Image_5	54/456	11.84 %	375/456	82.23 %
Image_6	35/412	08.49 %	345/412	83.73 %
Image_7	47/491	09.57 %	409/491	83.29 %
Image_8	54/374	14.43 %	276/374	74.60 %
Image_9	17/363	04.68 %	316/363	87.05 %
Image_10	26/320	05.50 %	280/320	87.50 %

 

 The fig 2 below shows the results of the Difference of Gaussian algorithm on four different partial fingerprint images. The extracted pores using DoG algorithms is marked with red circles.

 

Fig.2: Results of Difference of Gaussian pore extraction; (a),(c),(e) and (g) are input images;  (b), (d), (f) and (h)  shows detected pores on input images.

Here from the experimental results we can see that the true detection rates of the two pore models are more than sufficient. Also the false detection rate is low. The numbers of pores extracted from the two pore models are sufficient for reliable matching of pores. As seen from the Fig1 the number of pore are more than the average number of minutiae which testified our claim of higher feature count than minutiae count including ridge ending and bifurcations. Thus the proposed Dynamic Anisotropic Pore Model and Difference of Gaussian model has yield the sufficient features i.e. pores which when used with direct pore matching method will overcome the traditional fingerprint recognition technique using  minutiae   features.