Fingerprint Image Enhancement Using a Parallel ... - Semantic Scholar

3 downloads 0 Views 397KB Size Report
nakamura.takahiro@wrc.melco.co.jp hirooka.miwako@wrc.melco.co.jp fujiwahi@ina.melco.co.jp [email protected]. Abstract. Fingerprint image ...
Fingerprint Image Enhancement using a Parallel Ridge Filter Takahiro NAKAMURA Mitsubishi Electric Corp.

Miwako HIROOKA

Hideto FUJIWARA

Mitsubishi Electric Corp. Mitsubishi Electric Corp.

Kazuhiko SUMI Kyoto University

[email protected] [email protected] [email protected] [email protected]

Abstract Fingerprint image enhancement is an important process for improving the matching performance of low quality fingerprints. In order to eliminate high contrast noise lines such as deep wrinkles, which are prevalent in the low quality fingerprint images, we focus on the fact that true ridges have multiple paralleling neighbors. In this report, we propose the parallel ridge filtering method which can strongly suppress non-parallel noise lines by utilizing the parallelism of ridges. We also show the efficiency of our method by experimental results.

1. Introduction Noise in fingerprint images can derive from a variety of different conditions: creases cutting across fingerprints in the form of wrinkles or injuries, poor contrast as a result of dryness of the finger surface, indistinct ridge pattern from overall roughness of the finger, and so on. To eliminate these noise and restore the ridge pattern, usually some kind of ridge enhancement preprocessing is done before the actual fingerprint matching. However, the results of this preprocessing greatly affect the quality and the ability to make a match. In this work we focus on the inherent parallelism of true ridges (and furrows) in fingerprints and propose a parallel ridge filter that effectively screens out noiseinduced non-parallel lines as a new and more effective approach to ridge enhancement processing. We also demonstrate with experiments that our parallel ridge filter method works far better than the conventional approach in dealing with poor-quality fingerprints and yields significantly improved fingerprint matching performance.

2. Conventional Ridge Enhancement Rather than extract a ridge pattern directly from an image, most ridge enhancement processes seek to obtain a robust ridge pattern by some variant of the following two-step process: 1. The overall orientation of the ridge is derived.

2. By applying a specialized line extraction filter in the same direction as the orientation, the ridge pattern is reconstructed. In this approach, accurate extraction of the orientation is critically important. For example, Hong et al. have proposed a method based on the least squares averaging of the gradient of pixel values in a local window[1]. Various other approaches have also been proposed such as those described in Refs. [2] and [3]. Most of these conventional methods assume the following model with respect to fingerprints: A) Genuine ridges and furrows generally exhibit greater contrast than noise lines. B) Even if an error occurs in the orientation, the error can be corrected locally by using the continuity of the orientation with neighboring regions. Unfortunately, there are many regions in poor-quality fingerprints where these assumptions simply do not hold. Then the error in orientation becomes so extensive that it cannot be corrected, and the low-contrast ridges/furrows surrounded by high-contrast noise cannot be extracted. In contrast to the conventional approach built on the assumption of fingerprints containing little noise, in the next section we will consider a new model for relatively poor-quality fingerprints.

3. Parallel Ridge Fingerprint Model In this section, we propose a new fingerprint model that we shall refer to as parallel ridge model. The basic distinguishing characteristics of the model are that it discards all assumptions regarding contrast and focuses on the inherent parallelism of fingerprint lines. Let us first lay out our basic assumptions. A) A genuine ridge or furrow has more than a certain number of lines paralleling it on both or one side. Regarding other portions of the fingerprint not covered by the above we assume: B) In the vicinity of ridge (or furrow) bifurcations and endings (minutiae), the range where parallelism diminishes is so small. (At most, the radius is no more than one interval of ridges.)

0-7695-2128-2/04 $20.00 (C) 2004 IEEE

Figure 1 presents some examples of the lines extracted as the ones with enough number of paralleling neighbors by a preliminary experiment. We have found out that many minutiae with the lines bifurcating at a small angle fortunately satisfy assumption A), and although the lines tend to be chipped as the bifurcation angle increases, assumption B) still holds in almost all cases. So an overall orientation can be easily determined from the local orientations of the extracted lines by using a simple interpolation scheme. Furthermore, we experimentally estimate the area of the regions those are not covered by this model using a database of poor-quality fingerprints. Figure 2 shows the area rate of the regions where no lines having multiple paralleling neighbors are included within the radius of a ridge interval. Typical error regions are where no or few ridges exist according to roughness of skin or where ridge pattern is complex, but the area of such regions is limited. For comparison, we also show the area rate of the fingerprint regions those conventional scheme proposed in Refs.[1] cannot extract correct ridge patterns. One can see that the proposed model fits to poor-quality fingerprints far better than the conventional model. In the next section we shall describe ridge enhancement processing based on this model. small angle bifurcation

large angle minutiae and minutiae on bifurcation wrinkles rough skin

fingerprint image

Figure 1. Lines with at least 3 neighboring lines of which the orientation error is within 10 degrees (The utilized algorithm is to be mentioned in §4.2.)

㪉㪉㩼

㪉㪇㩼

㪈㪏㩼

㪈㪍㩼

㪈㪋㩼

㪈㪉㩼

㪈㪇㩼

㪏㩼

㪍㩼

㪋㩼

frequency

parallel ridge model constrast-dependent model (Refs.[1]) 㪉㩼

The ridge enhancement processing proposed in this paper is based on the following concepts. 1) First, we extract all lines appearing as if they could be ridges or furrows (candidate lines) without any attempt to distinguish between genuine lines and noise lines. 2) A candidate line is determined to be genuine if it is accompanied by a sufficient number of other candidate lines that are parallel to it; otherwise, the line is regarded as spurious. 3) The correct orientation is determined from candidate lines found to be genuine, and the ridge pattern is restored using a filter that extracts only lines paralleling the correct orientation. Now let us examine these processes in greater detail.

4.1. Extracting Candidate Lines As a filter that can extract clear intensity lines even under low-contrast conditions, here we employ a twodimensional Gabor filter cos wave kernel, namely: …(1) HT (x,y) = gauss(x’,y’) cos(2Sfx’) x’ = x cosT + y sinT, y’ = -x sinT + y cosT …(2)җ 1 1 x y exp(- {( ) 2+ ( ) 2}) ...(3) gauss(x,y) = 2 Vx Vy 2SVxVy where f is the frequency, and T is the filter rotation angle. The filter output in T directions for input image I(x,y) is given by

paralleling furrows (white pixels)

㪇㩼

4. Ridge Enhancement Using a Parallel Ridge Filter

(area without extracted ridge pattern) / (whole fingerprint area)

Figure 2. The frequency distribution of the area rate of the regions which do not fit to fingerprint models (The utilized database is to be mentioned in §5.)

GT (x,y) = I (x,y) * HT (x,y) (* represents the convolution integral)

 4)

We obtain the following candidate line image {TC(x,y), C(x,y) } using the direction that yields the greatest value. C(x,y) = Max{ GTi(x,y); i=1,2,…N } 

TC(x,y) = {Tҏi that gives C(x,y)}

…(5) …(6)

Where N is the number of orientation segments, and C and TC represent the intensity and orientation of a candidate line, respectively.

4.2. Determining Parallel Ridges Now we shall describe one method to determine genuine lines and suppresses the intensity of noise lines. First we select neighboring candidate lines Ci (i=1,2,…,M) with respect to candidate line C0 existing at point (x,y) by some method such as the one shown below. As illustrated in Fig. 3, M number of windows Wi (i = 1, 2, …, M) are set in a direction perpendicular to candidate

0-7695-2128-2/04 $20.00 (C) 2004 IEEE

line point (x,y). Then the single most intense candidate line satisfying the condition of the following equation is selected for each window, and defined as Ci. Max({C(a,b)}(a,b) Wi(x,y)) if |TC(a,b) -TC0|