Research on a fingerprint preprocessing method based on directional image

Abstract: A fingerprint preprocessing method based on directional map is proposed. The directional information of fingerprint image is used to realize fingerprint enhancement, binarization and extraction of irrecoverable area, which provides a feasible method for automatic fingerprint recognition.

Fingerprint features are one of the characteristics that remain unchanged throughout a person's life, and the probability that different people have the same fingerprint features is almost zero, so countries around the world are competing to research and develop practical fingerprint recognition systems. Fingerprint recognition systems generally consist of the following parts: fingerprint collection, preprocessing, feature extraction, classification and matching. In the process of fingerprint collection, various noises will inevitably be introduced, such as cross-connections and breakpoints in the image. These noises have a certain impact on the extraction of fingerprint feature information and may even produce many pseudo feature points. Therefore, before extracting fingerprint features, the fingerprint image needs to be filtered to remove useless information and enhance useful information. After obtaining the enhanced grayscale image, it needs to be further binarized to facilitate subsequent processing.

After studying many traditional grayscale image filtering algorithms, it is found that Fourier transform filtering has the best effect, but it is far inferior to the filtering effect of directional pattern filtering. In addition, the running time of Fourier transform filtering algorithm is very long. It takes several minutes to process a 512×512 image, while directional pattern filtering can be completed in just a few seconds. As for other filtering algorithms, such as median filtering and mean filtering, the effects are far inferior to directional pattern filtering. Therefore, whether considering the effect or speed, directional pattern filtering is a good grayscale image filtering algorithm.

Among the many filtering algorithms based on directional maps, one is to use various filters to filter after calculating the directional map of the fingerprint image using the direction of each point [1][2]. Another method is to use fingerprint ridge segmentation to achieve fingerprint enhancement. By using structural information such as local ridge direction and ridge width, a non-traditional binarization method is used to segment the ridge area and valley area from the original fingerprint image and represent them with a binary image [3][4]. This paper improves and supplements the second method by adding two processes: Gaussian filtering of each point and extraction of irrecoverable areas during the calculation process. Experimental results show that the effect achieved by this method is more ideal and reliable than the traditional second method.

1 Image normalization and extraction of fingerprint effective area

1.1 Fingerprint image normalization processing

Due to the characteristics of the collector itself and the structure of the finger, as well as the uneven force when collecting fingerprints, it is easy to cause the signal in some areas of the image to be too weak (too light in color) or too strong (too dark in color), which brings great difficulties to the subsequent fingerprint processing. Therefore, the fingerprint must be normalized so that the grayscale mean and variance of the lines in the image are close to the given expected mean M0 and expected variance VAR0. In this method, M0 and VAR0 are both 125. Grayscale image normalization does not change the clarity of the fingerprint texture.

Assume that the image I is of size N×N, let G(i, j) be the gray value of pixel (i, j), M and VAR be the image gray mean and variance respectively, G′(i, j) be the gray mean of pixel (i, j) after normalization, and the normalization process is shown in the following formula:

1.2 Extraction of effective fingerprint area

Since there is no change in the peaks and valleys of the ridges in the non-fingerprint area, its variance is very small, so the image is divided into multiple non-overlapping small squares of W×W, and the effective fingerprint area is extracted using the grayscale mean K and variance V of the small squares:

Where (i0, j0) is the coordinate of the upper left pixel in the block. For the obtained V, a certain threshold T1 is set. If V>T1, the block is a valid fingerprint area; otherwise, the block is an invalid area.

After the above operation, the fingerprint image is divided into fingerprint area and non-fingerprint area. Considering the connectivity of fingerprint area and non-fingerprint area in the fingerprint image, further processing is required, that is, removing isolated fingerprint blocks in large non-fingerprint areas and isolated non-fingerprint blocks in large fingerprint areas. After the processing is completed, the fingerprint image is marked as connected fingerprint area and non-fingerprint area. The non-fingerprint area is not within the processing range, and the fingerprint area needs further segmentation processing.

[page]

2 Fingerprint Image Segmentation

2.1 Filtering and Binarization of Fingerprint Images

In order to eliminate interference and enhance the lines, a line filtering enhancement method based on the directional map is used in view of the strong directional characteristics of fingerprint lines. To estimate the direction field, the direction of the fingerprint ridges is divided into 8 directions. The 8 fingerprint ridge directions at 1 pixel are shown in Figure 1. The angle between the directions is π/8, represented by 0 to 7. When obtaining the directional map of each point, since the image will be affected by various random noises during the acquisition process, the gray value G′(i, j) of the calculated point is not used directly. Instead, the rotational symmetry of the Gaussian low-pass filter is used to convolve the point set Ω formed by the point and the points in its eight neighborhoods (C1 to C8) with the Gaussian low-pass filter Gu to obtain the new gray value G″(i, j) of the point.

At the same time, calculate the sum of the grayscale values ​​of the pixels in the 8 directions of the point to obtain S0, ... S7. Divide the grayscale values ​​of the 8 directions into 4 groups (0-4, 1-5, 2-6, 3-7) according to the directions that are perpendicular to each other, and take the direction p with the largest difference between the two directions as the possible direction of the pixel point. The direction in which the grayscale average value is closest to the pixel value G″(i, j) of the point in the 2 directions is taken as the ridge direction at the pixel.

In the process of calculating the direction of each point, the ridge enhancement and binarization can be performed at the same time. If the pixel point is a point on the ridge line, the gray value of the point will definitely be greater than the gray average of all points in the 8 directions, and the average of the gray sum of all points in the ridge direction and perpendicular to the ridge direction will definitely be greater than the gray average of all points in the 8 directions. Therefore, combining the above two conditions will achieve better results.

[page]

If the current point C satisfies the above formula, then point C is on the ridge line. Set the grayscale value of point C to 125, otherwise point C is a background point and its grayscale value is set to 0. According to the above method, windows of sizes such as 13×13 and 17×17 can also be designed. The larger the window, the stronger the connection function for fingerprint breaks, but the isolation function for fingerprint adhesion is weakened, and the amount of calculation increases accordingly. Using the above method, the same operation is performed on all points in the image. After the operation, the breakpoints and cross connections in the image are well removed, and the initial segmentation of the image is also completed.

2.2 Extraction of irrecoverable areas

After successful fingerprint enhancement and preliminary segmentation, there may still be a part of the fingerprint valid area where the lines are very blurred, the lines are severely adhered or granular, and even the human eye cannot see the line structure, and the line structure in this area cannot be judged based on the line structure in the adjacent area. This area is called an irrecoverable area. It is necessary to further identify this area, and in the subsequent feature extraction process, feature values ​​are no longer extracted from the irrecoverable area, thereby avoiding the extraction of a large number of false detail points from it and improving the speed of detail point extraction.

In view of the fact that the directions of each pixel in the irrecoverable area are evenly distributed, while the directions of the normal fingerprint area are relatively consistent, the direction values ​​of each pixel calculated are used to extract the irrecoverable area:

The point direction map is divided into non-overlapping 16×16 blocks, and the direction consistency Ax of each block is calculated. The calculation process is as follows:

(1) Quantize the eight directions in Figure 1 into specific values. 0 equals 0, 1 equals -π/8, 2 equals -π/4, 3 equals -3π/8, 4 equals -π/2, 5 equals 3π/8, 6 equals π/4, and 7 equals π/8.

(2) When the directions of all pixels in each block are roughly the same, the absolute value of the sum of the directions of all points should be equal to the sum of all absolute values, and Ax = 1; when the directions of all pixels are evenly distributed, Ax = 0. Calculate the consistency of directions in each block and set a certain threshold T2. If Ax2, the area is set as an unrecoverable area.

3 Experimental Results

The above method was implemented on a microcomputer using Delphi programming. A fingerprint collector with a resolution of 500dpi was used to collect images of size 512×512. According to the characteristics of the collected image, the experimental parameters are as follows: the image enhancement area is 16×16, M0 and VAR0 are both 125; the block size in the effective area is 16×16, and the threshold T1 is 20; a 9×9 neighborhood size is used in the direction map; a 3×3 neighborhood is used in the Gaussian low-pass filter, σ＝1; in the irrecoverable area, the threshold of the direction consistency is T2=0.35; the experimental results are shown in Figure 2. Figure 2(a) is the original fingerprint image, Figure 2(b) is the image after contrast enhancement and extraction of the effective area, Figure 2(c) is the fingerprint image realized by traditional fingerprint ridge segmentation, Figure 2(d) is the fingerprint image after fingerprint enhancement and segmentation realized by the method proposed in this paper, Figure 2(e) is the fingerprint image after extracting the irrecoverable area, and Figure 2(f) is the fingerprint image after refinement of Figure 2(d).

[page]

4 Conclusion

Through experimental comparison of 50 pairs of fingerprints, it was found that in the traditional feature extraction method, the proportion of fingerprint pseudo-feature points in all feature points is 12%~14%. In the improved method, the ability to separate the discontinuous connection and fork connection of the fingerprint lines is greatly enhanced. At the same time, by marking the irrecoverable area, the extraction of a large number of pseudo-feature points is avoided, and the proportion of pseudo-feature points is reduced to 7%~9%. For the remaining pseudo-feature points, the pseudo-feature points caused by various types of noise can be deleted separately by using the calculated characteristics such as the direction of each feature point and the distance between each feature point[5][6], and the retained feature point set is used as the set of true feature points.

References

1. Jie Mei, Ma Zheng. Filtering algorithm based on ridge fingerprint. Journal of Electronics, 2004; 32(1)

2 Huang Xianwu, Wang Jiajun. Preprocessing combination algorithm for fingerprint recognition. Computer Applications, 2002; 22 (10)

3 Lin Guoqing, Li Jianwei. Preprocessing of fingerprint images. Computer Engineering, 2002; 28(9)

4 Jain AK, Hong L.An Identity Authentication System Using Fingerprints.Procedings of IEEE, 1997;85(9)

5 Yin Yilong, Ning Xinbao. Improved fingerprint detail feature extraction algorithm. Journal of Image and Graphics, 2002; 7(12)

6 Luo X, Tian J.Knowledge Based Fingerprint Image Enhancement.In: 15th ICPR, Barcelona, ​​2000