Research on preprocessing algorithm of fingerprint image automatic identification system

    Abstract: Systematically expounds the government and algorithm of fingerprint image automatic identification system preprocessing, and analyzes and experiments various technologies in fingerprint image preprocessing. After comparing various preprocessing methods, a better fingerprint image preprocessing algorithm was proposed and verified in experiments. Keywords: fingerprint recognition, preprocessing, image enhancement, binarization, refinement
   

    Because fingerprints have lifelong stability and amazing specificity, they have been used in identity identification for a long time and are respected as "the first of physical evidences". However, manual comparison of fingerprints is inefficient and slow, and is increasingly unable to meet the needs of modern society. In the 1970s, computer fingerprint analysis using image processing and pattern recognition methods began to be used in the United States to replace manual comparison. In the late 1970s and early 1980s, relevant equipment was developed and put into operation.

    Our country has successively carried out work in this area in the past 10 years and has now entered the practical stage. The technology of this discipline is used in fingerprint analysis and has two aspects of work:

    (1) Fingerprint feature extraction and identification. This is a relatively mature job, because everyone has a fixed fingerprint type, and the total types of fingerprint lines are limited. Therefore, it is possible to preprocess fingerprints, extract features, and then use cluster analysis method to make judgment classifications.

    (2) Fingerprint storage and archiving. This is a problem of large-capacity storage. Just imagine that each person has 10 fingerprints, and more than one billion people in the country have to store tens of billions of fingerprints. This requires huge storage capacity. In addition to using high-density recording and storage methods such as laser storage, there is also an urgent need to carry out distortion-free and recoverable compression of fingerprints.

    In recent years, automatic identification systems using computers for fingerprint comparison have been developed abroad. Fingerprint images have a large amount of information and can be processed in a timely manner. These systems are based on medium and small computers or even dedicated high-speed parallel processors.

    In order to quickly carry out research on automatic fingerprint identification, we tried to develop a microcomputer automatic fingerprint identification system on a microcomputer. The input part of the system converts actual imprinted fingerprints or on-site fingerprints into digital images, which are stored in the image memory via the bus. The microcomputer performs various necessary processing and classification on the input images. The processing effect can be immediately evaluated by a color monitor or by a printer. Output the judgment results.

    The principle of a general fingerprint image automatic recognition system is shown in Figure 1.

    The purpose of fingerprint image preprocessing is to make the fingerprint image clear and have obvious edges so that features can be extracted for identification. The main purpose of preprocessing technology is to highlight certain information in the fingerprint image and weaken or remove certain unnecessary information for a given fingerprint, so that the result is more useful for subsequent identification than the original latitude and longitude image. suitable. A certain preprocessing flow chart is shown in Figure 2.

    Only pre-processed fingerprint images can proceed to the next step (fingerprint recognition and pre-classification) for processing. The quality of the preprocessing results is directly related to the fingerprint recognition rate.

1 Image enhancement

    In terms of processing methods, image enhancement processing can be divided into two types: spatial domain method and frequency domain method. The spatial domain method refers to direct processing in the spatial domain where the image is located; while the frequency domain method refers to performing Fourier transform on the image first, and then performing inverse Fourier transform after processing in the frequency domain.
We used two steps of smoothing first and then sharpening. This method can automatically process fingerprint images and is more effective.

1.1 Smoothing processing

    The fingerprint image is an analog signal obtained through the scanner, and after sampling and quantization, it is stored in the computer in the form of a matrix. Since the image is collected in a column mode, the quantized fingerprint image contains a lot of noise. Due to their temporal incoherence, they contain high spatial spectrum and mostly have point structures.

    The task of smoothing is to remove these interfering noises without distorting the image. Image smoothing includes two categories: spatial domain method and frequency domain method. According to the air domain processing formula, the output function is:

    where m=1,2...,H'; n=1,2,...N'; M'=ML; N'=NL. If the input image matrix F is N×N, the convolution matrix H is L×L, and the output matrix G is M×M, then the relationship is:

    In order to smooth the noise, the convolution matrix H needs to be low-pass. When H is a 3×3 array, choose:

    where X is the pixel value to be processed, and ai (i=1,2…,8) is the value of its eight neighbors.

    Using the multi-image averaging method (that is, the method of averaging multiple inputs) can remove temporal noise. It is based on the statistical characteristics of noise interference. That is, if an image contains noise, it can be assumed that the noise caused by this is uncorrelated with respect to each coordinate point (x, y). Its digital expectations are zero. Let g(x,y) be the superposition of noisy eta(x,y) and the original image f(x,y). That is:

    g(x,y)=f(x,y)+η(x,y) (4)

    If the noise η(x,y) satisfies the above assumption, then a given series of noise images { After superimposing gi(x,y), take the average value to achieve the purpose of smoothing the image.

    Therefore, when obtaining an image, the method of collecting M times can be used to obtain the average value, which can meet the requirements.

g(x,y)=1/M∑gi(x,y) (5)

    Because:

    E{g(x,y)}=f(x,y) (6)

    Generally, when n=4 , the image tends to be stable.

    Practice shows that using these methods to remove noise can achieve satisfactory results. Figure 3 shows the fingerprint image after noise removal.

1.2 Sharpening Processing

    In order to strengthen the boundaries between fingerprint lines and highlight edge information to facilitate binarization, sharpening processing is necessary. Sharpening is useful for enhancing contrast and detecting edges. The function of image sharpening is to compensate for the contours of the image to make the image clearer. Image sharpening can be divided into two types: spatial domain image sharpening method and spatial frequency domain image sharpening method.

    The blur of the image is caused by the fact that high spatial frequency components are weaker than low spatial frequency components. This effect is manifested in the boundaries (edges) between uniform grayscale areas. Therefore, to eliminate blurring, high spatial frequency components can be enhanced.

    As a simple high-frequency enhancement filter in image space:

    g(i,j)=f(i,j)-f(i,j)=5f(i,j)-[f(i+1,j)+ f(i-1,j)+f(i,f+1)+f(i,j-1)]

    where f(i,j) represents the numerical value of the second differential of the input image f(i,j) The Putian operator is defined as:

    f(i,f)=f(I+1,f)+f(i-1,j)+f(i,j+1)+f(i,j -1)-4f(i,j) (8)

    Since the blurred image is caused by the diffusion phenomenon during the photosensitivity process, it must satisfy the linear diffusion agenda:

    аf(x,y,t)/аt=kf(x ,y,t)=k(аf/аx2+аf/аy2) (9)

    Among them, f (x, y, t) is a function of time and space, k is the diffusion coefficient; f is the Laplacian operator of f. If t=0, and ignore the higher-order terms of its expanded Taylor series, the output function is:

    g(x,y)=f(x,y,t)-τkf(x,y,t) (10)

    where τ is the length of the diffusion time interval. By selecting different τk values, different unweighted coefficients of the Laplacian operator can be obtained. After many experiments, it is known that the weighting coefficient p of the 7×7 square matrix is:

    p=8f(x,y)-f(x-3,y)-f(x-2,y+2)-f(x, y+3)-f(x,y-3)-f(x+2,y+2)-f(x+3,y)-f(x+2,y-2)-f(x-2 ,y-2) (11)

    can better reflect the continuity of the function during Taylor series expansion, and is suitable for sharpening processing of fingerprint images. Figure 4 is the sharpened fingerprint image.

2 Binarization processing

    Binary images refer to images with only black (grayscale value is 0) and white (grayscale value is 1) binary images in the entire image, and there is no change in grayscale on them. In digital image processing, binary images occupy a very important position. This is because in a practical image processing system, high processing speed and low cost are required, and shading image processing with a large amount of information is too expensive and is not a good idea. Moreover, the binarized image can be analyzed and characterized using concepts in geometry, which is much more convenient than grayscale images. Therefore, binary image processing has now become an independent and important branch in image processing and has been widely used. For fingerprint recognition, some information must be included in the binary description of ridges and valleys. Therefore, it must be determined based on the original grayscale image whether each point on the image belongs to the object area or the background area, thereby generating a corresponding binary image. It can not only greatly reduce the storage amount, but also make the subsequent identification process less disturbed and greatly simplify the subsequent processing method. The binarized image is the basis for subsequent processing, and its algorithm has a direct impact on subsequent processing. A good algorithm can obtain a high-quality binary image. On the contrary, if noise is introduced at this stage, it will directly reduce the image quality and affect the recognition accuracy.

    The threshold V(x,y) of the pixel f(x,y) located at (x,y) is determined by the points in the fingerprint image window (2m+1)×(2m+1) centered on (x,y) determined by the grayscale value. That is:

    when the (2m+1)×(2m+1) window falls in two different areas, it will cause misjudgments:

    (1) When the window is mostly in the valley line area, some or even most of the pixels will be Points are judged as ridge lines;

    (2) When the window falls more in the ridge area, some or even most of the pixel points will be judged as valley lines.

    This problem can be solved with the help of parameter β and fixed threshold.

    Let ε be a positive integer obtained through experiments, that is, ε>0, so there is:

    where Vt is the reference threshold.

    Let the binarized image be g(x,y), then:

    the above process can be represented by flow chart 5.

3 Modification processing

    After the fingerprint image is binarized, due to quantization and other reasons, the edges of the lines are uneven. Affected by sharpening, discrete points appear on the screen. In order to make the image neat and smooth, modification processing is required.

    Discrete black spots in the grooves can be removed and empty white spots in the grain lines can be filled with a suitable template. For example, let the pixel to be processed in the 3×3 template be x, and its neighborhood ai∈white (i=1,2,…,8), then x=white; ai∈black (i=2,4,6,8 ), then x=black.

    Modification processing includes operations such as removing isolated points and filling blank points. Their respective templates are represented as follows:
  
  4. Refinement processing

    Since what we are concerned about is not the thickness of the lines, but the presence or absence of the lines. Therefore, redundant information must be removed without destroying image connectivity. In the field of binary image processing, thinning is a very important processing step. For example, in an interference fringe image, since the fringes are thick and wide and the edges are diffuse, there is no way to accurately calculate them without refining them into lines. On the other hand, if a connected component can use a linear structure to refine the image, it has become a hot topic in image processing. There are many contours of human pattern ridges and related endpoints and intersections. However, the binarized ridge has a certain width, and this characteristic that characterizes it is of great significance in both image recognition and data compression. Therefore, the second width will add a lot of trouble to subsequent identification. Therefore, the width of the fingerprint ridge should be gradually peeled off first, so that the ridge becomes a thin line only one pixel wide, which will be very helpful for the next step of analysis. This process is called thinning. The purpose is to use a set of thin lines to depict a connected component, which not only achieves the purpose of compressing the amount of data, but also facilitates feature extraction of the connected component. Therefore, this operation necessarily requires the deletion of some pixels.

    The algorithm for thinning using two types of templates is as follows:

    Figure 6 shows three different types of fingerprints obtained after processing by algorithm (19).

    After the fingerprint image is refined, general processing rules can also be obtained based on the fingerprint. Perform post-processing, such as removing invalid line bridges and connection gaps, etc. Human-computer interaction can also be used to manually repair incomplete fingerprints.

    This article provides a complete set of fingerprint image preprocessing algorithms, and achieved satisfactory results in experiments.