The last few decades have witnessed exponential rise in advanced technologies. Despite significant innovation and technological horizon, personalized security often remains a challenge under dynamic application environment. Unlike cryptographic concepts, in the last few years biometric driven authentication systems have increased significantly
Considering above stated issues and scopes
Finally, assessment of both the proposed methods is carried out by obtaining Confusion Matrix for both the methods and concluded that textural features help in improving the efficacy of the classification of contactless 3D fingerprint classification. Also, the proposed method has reduced Equal Error Rate when compared with existing state of the art methods.
This section primarily discusses the overall proposed contactless fingerprint detection and classification system. Unlike classical fingerprint detection models, in this research the focus is made on improving input data environment as well as feature vectors to
In sync with the targeted contactless environment for fingerprint detection system, in this work we collected contactless three-dimensional sensor driven images to prepare datasets. The 3D contactless fingerprint datasets were collected in such a manner that it could enable effective learning under data heterogeneity and diversity to make it more efficient under realistic environment. We considered the 3D Fingerprint dataset comprising a large contactless fingerprint sample. Noticeably, for our case study we considered a total of 50 subjects and the samples collected were from the subjects aged in between 28 to 55 years. The subjects comprised a total of 40 man and 10 women that eventually contributed 160 and 40 fingerprint samples, correspondingly. The data considered had been collected under natural light conditions with standard illumination. Here, no specific light or illumination control measure was applied. Also, for comparison with the state-of-the-art methods, In the proposed method, data samples from 1000 random users are collected from different benchmark databases (Hong Kong Polytechnic University 3D-fngerprint images Database Version 2.0, IIT Bombay, IIT Kanpur, Touchless Fingerprint Database, UNFIT database from Image Analysis and Biometrics Lab. IIT Jodhpur) are considered.
As already stated, realizing practical Touchless input acquisition and allied complexities, we processed each input sample for pre-processing. To achieve it, at first the input images were processed for image resizing. In this method, firstly the centroid of the fingerprint image was spotted using region –property function. With spotted centroid as a reference a radius of 120 pixels is marked and a circular ROI was estimated for each image. Once getting the ROI, it was handled for RGB to GRAY transformation. Over Gray output, histogram equalization was performed that alleviates major key problem of intensity variation over the retrieved images and makes it suitable for further processing. Later we have performed normalization followed by binarization. After obtaining the binary image of the fingerprint, we have performed thinning function that enabled skeleton formation of the image.
Above mentioned pre-processing tasks helps in extracting most of the ridge and structural information to achieve better accuracy. This process also helps in redundant data elimination that results in improvement in efficiency. The detailed discussion of the proposed minutiae estimation model is given in the sub-sequent section.
Typically, a justifiable illustration of a fingerprint can be provided in terms of its corresponding minutiae details. Among the all-possible minutiae details the ridge endings and bifurcations are the dominant one. Usually, ridge endings signify the points where the ridge curve stops or terminates, while bifurcations signify the specific location where a ridge splits from a single path into two paths at certain Y-type junction. To extract minutiae, we applied a simple Crossing-Number concept. In our applied crossing number concept, the binary image where the ridge flow pattern is eight-connected is considered. Now, performing scanning of the local neighborhood of each detected ridge pixel in the binary image with a
In above equation, the variable
SSIM is one of the most used image analysis tools used for quality assessment and image-matching purposes. It exhibits the similarity of the two distinct images based on certain features such as structural distortions, textural features, luminance etc. It performs image comparison not based on the pixel values rather the image elements identified by human. SSIM embodies varied distortions, including contrast, luminance and texture to compare fingerprint images. Noticeably, in our case, we applied minutiae details as the feature in SSIM to match fingerprint images of the users to make identification decision. The suitable values of SSIM can be in the range of -1 (maximum difference) to 1 (no difference). For our considered fingerprint minutiae picture, SSIM at coordinate
and when
The definitions of the different parameters applied in above equation are given as follows:
α, β, γ – Significance coefficients
Thus, obtaining the value of (8), the two images can be same only when it has the SSIM value of 1. On contrary, low SSIM value signifies the disparity between fingerprint images or unmatched fingerprint.
Let,
Consider that the orientation image be O, with the dimension
In this proposed work, GLCM functions as a descriptive statistical feature distribution model assessing the probability of the pixel’s gray scale values over an input fingerprint image. Functionally, it extracts high-dimensional statistical features. In this work, we assume that after preprocessing, the varied textural features are distributed uniformly throughout the pre-processed input image. In this reference, over each input fingerprint image we extracted the different textural features, which were later combined together to yield a composite feature vector for classification. In this method, the retrieved textural features were derived in the form of a matrix representing pixel intensities
With the extracted values of
Contrast,
Energy,
Homogeneity,
Correlation,
Mean,
Standard deviation,
Variance,
Kurtosis, and
Skewness.
As stated, a total of nine STTF features were obtained for further feature learning. Here, our predominant goal was to retain maximum possible and significant features for learning and classification so as to achieve higher accuracy. The brief of GLCM features extracted is discussed below:
In GLCM, contrast is defined as the variation in gray scale parameter values over the input image. With the derived probability matrix (11), the pixel pairs representing the diagonal element represents vital difference in contrast values. Here, the texture contrast represents the cumulative variations in the local pixel intensities across the input fingerprint image. Generally, the non-linearity existing across the input image is examined by performing statistical estimation and corresponding textural continuity assessment. We measured the contrast information for each input fingerprint image using equation (16).
To examine energy distribution across the image, the proposed model measured angular second moment (ASM) value that measures the rotational acceleration over the input feature space. Here, the model defined in (12) was applied to estimate the ASM value. Typically, the value of ASM increases linearly throughout the gray-level values over input image.
Once calculating the ASM values (12), we estimated the energy parameter using (13).
Entropy, which is also referred as the pixel-disturbances within an input image. In other words, it also signifies how non-linear the gray-level values are distributed throughput the input fingerprint image. Typically, the entropy value increases with rise in pixel’s non-linear distribution. We applied (14) to estimate the entropy value.
Typically, homogeneity refers the Inverse Different Moment (IDM) signifying higher homogeneity with reference to the lower contrast. In other words, an input image would have lower homogeneity with higher contrast. In this work, we used equation (15) to measure homogeneity distribution throughput the input fingerprint image.
In reference to the linear magnitude distribution over input fingerprint image, smaller value of contrast would give rise to the higher homogeneity. In this work, we employed equation (16) to measure contrast over the input image.
Correlation signifies the statistical feature representing descriptive statistics throughout the input image. In this work, in addition to the correlation information (16), we extracted three other spatial-temporal statistical features encompassing mean, standard deviation, and variance. In order to estimate the mean value, we employed the symmetric features of the probability matrix
Further, we applied mean values (17-18), variance and standard deviation values were obtained by applying equations (19) and (20), respectively.
Thus, estimating the mean and variance values, we estimated correlation information using (21).
With the above stated statistical features, we extracted directional or orientational features, including skewness and Kurtosis. These features are also called as the symmetrical statistical features. With the estimated probability matrix (11), skewness parameter refers the lack of symmetry. In case of at image processing tasks, skewness is defined in the form of shade feature where the high cluster-shade signifies asymmetrical nature. We applied equation (22) to estimate the skewness values per input fingerprint image.
Kurtosis signifies strength of the input gray-level values distributed throughput the input fingerprint image. Here, higher Kurtosis confirms or indicates that the amount of the feature-distribution is mainly strenuous along tail than the mean value. The lower value of Kurtosis indicates that the amount of feature-distribution remains strenuous in the direction of the spike which is closer to the mean value. In this work, Kurtosis was obtained over the entire input image as there is no specific target area and throughout image serves as an input textural feature to make learning and further classification. Once extracting above stated nine different GLCM features, we performed horizontal concatenation to estimate a composite feature vector for further learning. The composite GLCM feature obtained is given in equation (23).
Now, once estimating the composite feature vector (i.e.,
Random Forest is one of the most successful learning algorithms that structurally encompasses multiple tree-based classifiers, behaving as an ensemble learning model. In the proposed tree-model, each tree provides its corresponding vote for the most probable class for each input fingerprint image. Let the total training samples be N, then a sample encompassing N cases are randomly selected from the original data. These selected samples are further employed as training set to form a new tree. Now, in case there are M input variables, then the best split on these M is applied to split the node. Here, we maintained the value of M as constant during forest development, also called as the growing phase. In this manner, each tree is developed to the largest extent. Unlike classical machine learning methods, Random Forest algorithm needs smaller number of parameters to be estimated during classification. It makes overall computation more efficient and suitable for the real-time uses. A complete Random Forest algorithm can be eventually defined as the combination of the different tree-structures, as presented in (24).
In (24), the parameter h signifies the classifier function, while
Considering the significance of a touchless 3D fingerprint identification and classification system, in this research as a first method a minutiae-based measure was proposed where we focused on at first setting up an optimal data environment, followed by feature extraction mechanism and pattern matching followed by classification. Understanding the way that the touchless data acquisition might undergo different local environmental changes such as change in brightness, contrast and skin defects an efficient pre-processing was carried out and the results of preprocessing are shown in
Now, we mainly focus on assessing efficacy of the proposed textural feature driven contactless fingerprint detection and classification model, qualitatively as well as quantitatively. In other words, here we examine whether the use of local pre-conditioned image improvement yields superior performance. Before discussing the simulation results quantitatively, a snippet of pre-conditioned and enhanced results is shown in
|
|
|
|
|
X. Yin, Y. Zhu, and J. Hu |
3D Reconstruction |
3D Stereo images |
TTP features |
0.66 |
J. Galbally. L. Beslay G. Böstrom |
Deep CNN |
3D-FLARE DB |
HOG+LBP |
1.04 |
Bakheet, S, Alsubai, S, Alqahtani, A, Binbusayyis |
SIFT |
FVC2004 |
Minutiae |
2.01 |
Priesnitz, J, Huesmann, R, Rathgeb C, Buchmann N, Busch |
VeriFinger |
PolyU |
Minutiae |
3.17 |
Attrish A, Bharat N, Anand V, Kanhangad V |
Deep CNN |
IITI-CFD |
Minutiae |
2.19 |
Birajadar P, Haria M, Kulkarni P |
VeriFinger |
IITB |
Minutiae |
1.18 |
|
GLCM + RF |
PolyU+ IITB |
Textural Features |
|
This is the matter of fact that a large number of studies have been done towards touch-based fingerprint detection systems; however, the efforts made towards touchless fingerprint detection are countable and very rare. Our depth literature assessment revealed that merely countable a dozen of efforts is made so far to introduce 3D touchless data for fingerprint detection. To assess relative performance, we have selected the couple of recent methods listed in
The fingerprint detection models have always been considered as a vital alternative of the classical cryptosystems. Undeniably, being fast in execution fingerprint-based systems turn out to be more efficient solution for personalized security. This efficacy makes fingerprint-based authentication system as one of the most used approaches. Despite robustness, being touch-based paradigm, its optimality has challenges under different operating environment, especially in reference to the health and hygiene. During the recent pandemic of COVID-19, touch-based fingerprint models were found vulnerable due to touch-based infection probability. To alleviate such issues, contactless fingerprint detection method can be a viable solution; however, being touchless in nature such approaches might undergo different complexities like the impact of viewing angle, textural non-linearity, non-uniform illumination and contrast, ridge and furrow ambiguity, ridge discontinuity etc. Extracting conventional structural features like minutiae over aforesaid local adversaries can impact overall efficacy. On the other hand, to cope up with touchless environment demands, an improvement in local conditions and feature modalities followed by training over textural features improves accuracy and also minimizes EER. Quantitatively the Proposed method outperforms when compared with the existing state of the art methods on the benchmark database by achieving an accuracy of 94.72%, precision of 98.84%, recall of 97.716%, F-Measure 0.9827 and a reduced EER of about 0.084. The key novelty of this approach was that it didn’t require any surface reconstruction, rather it employed different mathematical approaches to retrieve surface normal and minutiae information. As a future work we can experiment with the different feature extraction methods like CNN or hybrid techniques also we can experiment with ensemble learning algorithms to still improve the accuracy of classification.
The Authors would like to thank the management, Principal and authorities of Malnad College of Engineering, Hassan for extending full support in carrying out this research work.