Improving of fingerprint segmentation images based on K-means and DBSCAN clustering

ABSTRACT


INTRODUCTION
In present world, fingerprints have become an important biometric technology due to its uniqueness and invariance to every person. Moreover, with the popularity of fingerprinting technology, especially in mobile phones, the technology previously used in the field of criminal investigation is now commercialized. This biometric trait is more used and acceptable by the users because the capturing device is relatively small and identification accuracy is comparatively very high to other biometric recognition techniques such as the retina, iris, hand geometry, etc. [1][2].
The image segmentation is one of the main problems in the field of computer vision and image processing. Therefore, the fingerprint segmentation is typically the first and foremost step in the process of biometric recognition system based on fingerprint. In addition, the effect of this step directly affects the performance of the system.
The general structure of fingerprint recognition system consists of four major steps. In the first one, the acquisition of fingerprint image is process of getting a digitalised image of a person by using the specific sensors. These images can be acquired in two ways: offline and live-scan acquisition [3][4]. In the second

RELATED WORKS
The fingerprint image segmentation is one of the principal stage for automated fingerprint recognition system. This pre-processing stage allows to separate the fingerprint region from a captured image with two areas: foreground and background [7]. Most existing techniques of segmentation are based on the feature of pixel intensity in a block because it is computationally faster than others based only on the pixel intensity [8][9][10]. For fingerprint segmentation, there are so various methods have been proposed in the stateof-the-art. Here, we briefly review these methods.
Li, et al. [11] proposed a segmentation technique by calculating gray contract and Fourier spectrum energy ratio for each block in fingerprint image and then classified these block by linear support vector machine approach. Finally, morphological operations are used to improve the segmented image. Akram, et al. presented a segmentation method by computing mean, variance and gradient deviation information of each block in fingerprint image. The segmentation image of fingerprint is obtained by using the linear classifier [12]. Li, et al. [13] suggested a method for fingerprint image segmentation using a novel approach of K-Means. the fingerprint image is divided into non-overlapping blocks. Furthermore, for each block, the variance, direction and energy spectrum are extracted to construct feature vectors and then, classified these characteristics by K-means clustering algorithm. Finally, used the post-processing to remove the remaining isolated blocks in foreground or background region. In Yang, et al. [14][15] have been subjected a novel algorithm of fingerprint images segmentation by using an unsupervised learning method based on K-means classifier. Thus, for each block, the average and coherence data is computed in order to divide the image into two regions by using K-means clustering approach. The correlation based fingerprint image segmentation is used in [16]. Fahmy, et al. [17] proposed a technique that utilizes morphological processing to extract the foreground from the fingerprint image. After the division of image into nonoverlapping blocks, this method used the feature vector for each block to realize fingerprint segmentation. Then, the adaptive thresholding is used to convert the fingerprint image to a binary one. Next, some morphological operations (closing and opening) are applied, to segmented the image. Finally, the complex Fourier series expansion are performed to smooth the segmented contour. In this algorithm, the image is separated into blocks and sub-blocks. Afterwards, the thresholding level have been applied for segmentation. Das, et al. [18] achieved the fingerprint segmentation by computing block based statistics and morphological operations. Abboud, et al. [19] presented a new segmentation technique by statistical computing: mean, variance and coherence features of each block in fingerprint image based on an automatic threshold values and Otsu's method. Finally, the filing the gaps is applied to remove the noise in some regions in foreground or background by using particular sets of rules based on neighboring regions.

PROPOSED APPROACH
Our proposed method is improved the fingerprint image segmentation based on K-Means and DBSCAN clustering. A robust and effective fingerprint image segmentation algorithm is important phase for a fingerprint recognition system. In this section, we detail the proposed technique which is illustrated in Figure 1. The details of each phase are represented in the following.

Pre-processing
In this phase, Sobel and TopHat filter method have been used to improve the quality of the fingerprint image. Sobel structuring operators Sobelx and Sobely for image are represented in (1).
The gradient Gx and gradient Gy of pixels are defined from image Img by (2). G x = Sobel x * I mg , G y = Sobel y * I mg (2) The result of gradient is combined to find the absolute magnitude (the output edge). This result is described as follows: The fingerprint image is ameliorated after normalisation and Sobel technique. However, the fingerprint image is more improved by using the TopHat technique. This filter is a process that extracts details and small elements from image. TopHat filtering is based on dilation, erosion, opening and closing method. The morphological dilation and erosion operation for image Img of size x×y with structuring element Se are defined by (4) and (5) respectively: The opening and closing process for image Img with structuring element Se are described by combining the erosion and dilatation operation given by (6) and (7) respectively: The opening TopHatop and closing TopHatcl operations for image Img with structuring element Se are represented by (8) and (9)

Segmentation
After the pre-processing phase, the fingerprint image is divided into non-overlapping local blocks of size w×w. Further, to every block, the characteristic vector is classified into two classes: foreground and background region by using K-means classification.

Characteristics extraction
The characteristic vector is represented, for each block in fingerprint image, by its three categories namely: image intensity based characteristics, gradient based characteristics and ridge based characteristics.

a. Image intensity based characteristics
The change in intensity values is usually specific along the ridges and no-ridges when compared to background areas in fingerprint image. General image intensity based characteristics can be used to define the most intensity such as difference of mean, which is the difference between the local intensity mean and the global intensity mean, and variance blocks in given image Img of size x×y. These proprieties are computed by (11) and (12) respectively.
where ImgMeanL is the mean intensity of the local block of image and ImgMeanG is the mean intensity of the global image.

b. Gradient based characteristics
The gradient is utilized to obtain the directional variation in intensity value along a direction of image Img of size x×y. Characteristics such as gradient cohrence and ridge direction can be classified into gradient based characteristics. The gradient coherence and ridge direction features are calculated by (16) and (20) respectively.

Local variance thresholding
The variance local thresholding decides whether or not the region in fingerprint image needs to compute another characteristic for segmentation. The local variance is calculated in given image Img of size x×y by (12). If local intensity variance in a w×w image block is greater than 0, the other characteristic of the block will be computed otherwise are considered as null. Therefore, from Table 1, in DB1 it is observed that the difference of the processing time of our proposed segmentation is achieved almost 10 seconds for an obtained fingerprint image. Additionally, it is shown that the average segmentation time in 4 databases is less than other method based on K-means segmentation which the block is represented just by 3-dimensional characteristic vector [13].

K-means clustering
In our proposed method, K-means classifier is utilized to classify the five extracted characteristics, variance, difference of mean, gradient coherence, ridge direction and energy spectrum, from each local block in fingerprint image in order to distinct the foreground area to the noisy background area. The characteristic extraction vector for each block is represented by (23). On the one hand, the K-means algorithm is a popularly unsupervised machine learning models [20] used for clustering technique of the data [21]. On the other hand, it is simplified to implementation, eased of interpretation, faster and adapted to sparse data. K-means classification is represented as follows. Firstly, this approach selects randomly the number of clusters and assigns the cluster with closest centroids; then it determines each data point to the nearest centroids and for every cluster, the new centroids are recalculated, and this process is repeated until some condition is verified [22].

DBSCAN algorithm
DBSCAN, Density-Based Spatial Clustering of Applications with Noise, is another model of machine learning and clustering analysis algorithm proposed by Martin Ester et al. [23]. DBSCAN is one of the most commonly used cluster analysis algorithms. This algorithm is density-based: given a set of points in a dataset, the algorithm can the nearby points are grouped together (points with many adjacent points) and the outliers in the low density area are marked. The general algorithm of DBSCAN Clustering is shown in The reason of using DBSCAN over other clustering algorithms is that it does not demand a fixed number of groups in dataset. It also recognizes outliers as noise which simply introduces them into the cluster even if the data points are very different. In addition, it is a good place to find groups of any size and shape. In our propsed algorithm, DBSCAN Clustering is used to achieve more compact blocks for reducing the misclassification region due to the K-means algorithm . Figure 2 presents the result using the DBSCAN algorithm. Finaly, the contour smoothing (filtering in a complex Fourier transform domain [24]) as postprocessing technique to smoothen the edges of mask [17].

RESULTS AND DISCUSSION
The experimental operation platform in this study is described as follows: the host configuration: CPU Intel Core2 Duo at 2.00 GHz, RAM 3.00 GB, runtime environment: Microsoft Visual Studio C++ 2013 with OpenCV library. In order to better verify our algorithm, the following segmentation methods are adopted in the experiment: SVM [9], K-means with 3-dimension feature [13], MP [17], ACT [19]. These segmentation algorithms were compared to each other. In order tovalidate the proposed algorithm, the results have been tested on the public Fingerprint Verification Competition 2004 dataset [25] which contains 4 databases, namely DBl, DB2, DB3 and DB4. The performance measure is used the number of misclassification as defined by (25).
where Nbrbe is number of background classified error, Nbrb is the total number of true background regions in the fingerprint image and Prob1 is the probability that a foreground region is classified as background. Nbrfe is the number of foreground classified error, Nbrf is the total number of true foreground regions in the fingerprint image and Prob2 is the probability that a background region is classified as foreground. The probability of error ProbErr is the average of Prob1 and Prob2.
The existing algorithms of fingerprint image segmentation, SVM [11], K-Means with 3-dimension feature [13], MP [17] and ATC [19], are implemented for comparison of segmentation performance. The comparison be tween proposed method and others works in measure of average segmentation error for fingerprint images at different databases is shown in Table 2. From Table 2, in The SVM [11] and K-Means with 3-dimension feature [13], the corresponding value for misclassification rate in DB1 is 18,75%, 20,28% respectively. Furthermore, the error rate of segmentation in MP [17], ATC [19] and our proposed method is 0,28%,13,31% and 0,30% respectively. In the second database, if SVM [11], K-Means with 3-dimension feature [13], MP [17] and ATC [19] give 34,56%, 22,30%, 2,92%, 29,79% respectively in segmentation rate however our proposed method takes minimum error rate as 1,66%. Likewise, in the last database, misclassification rate (1,09%) is clearly less for the proposed algorithm compared to other techniques in the third database. SVM [11] has failed to classify better the forwound and background region (error rate as 28,89%) in the fourth database. thus, our system has succeeded to reduce 0,63% of segmentation error than K-Means with 3-dimensions features [13], MP [17] and ATC [19] which give 5,06%, 1,31% and 17,36% respectively. The proposed algorithm results show a better performance respect to other algorithms with average misclassification rate as 0,67% at different databases of fingerprint images. The means of these results shows that our algorithm is improved the recognition accurate rate of the person.From these existing works, it is worthy to say that the results of proposed method are clearly superior in term segmentation error rate. The Figure 3 represents the visual results of the proposed algorithm and the other techniques. When we compared to the existing works in this figure. We can say that our proposed is efficient and reduced error rate of segmentation.The visual quality results of segmented image indicate that the proposed algorithm adapts and gives better results in term of segmentation in different environments than the comparative techniques for accuracy segmentation.   [11], (c) K-means [13], (d) MP [17], (e) ATC [19], (f) Proposed method

CONCLUSION
This paper has proposed an improved method of fingerprint segmentation images based on K-Means classification and DBSCAN clustering. Our proposed system is presented in three steps. In the first step, the quality of fingerprint images is enhanced using Soble and TopHat filtering method. In the second step, K-means approach is applied for each local block to classify the image into foreground and background region using five-dimensional characteristic vector extraction. Moreover, the processing time of segmentation is faster than another algorithm based on just 3 dimensions features due to the local variance thresholding. In the final step, DBSCAN algorithm is used to remove some misclassified blocks due to the K-means clustering and contour smoothing is achieved to improve the image segmented. Simulation results show significantly that the proposed method is efficacy some recent existing techniques in average segmentation error rate. Therefore, it affects the performance of the system to have a higher accurate recognition rate of the person.