JOIV : Int.
Inform.
Visualization, 8.
: IT for Global Goals: Building a Sustainable Tomorrow - November 2024 1635-1642
INTERNATIONAL JOURNAL
ON INFORMATICS VISUALIZATION
INTERNATIONAL
JOURNAL ON
INFORMATICS
VISUALIZATION
journal homepage : w.
org/index.
php/joiv An Automated Fingerprint Image Detection and Localization Approach-based Unsupervised Learning Algorithms using Low-quality Biometrics Plam Data Abdulrasool Jadaan Abed a,*.
Dhahir Abdulhadi Abdullah a Department of Computer Science.
College of Sciences.
University of Diyala.
Iraq Corresponding author: *scicompms222315@uodiyala.
AbstractAiIn this study, fingerprint identification and classification of low-quality fingerprints have been analyzed accordingly.
technology advances and methodologies evolve, staying at the forefront of research and innovation is imperative.
The challenges addressed in this paper provide a foundation for future investigations and underscore the importance of developing resilient and adaptable biometric systems for real-world applications.
The quest for accurate, efficient, and robust fingerprint identification in adverse conditions is a testament to the continuous evolution and refinement of machine learning and deep learning approaches in While deep learning models exhibited improved performance, it is essential to acknowledge the need for further research and development in this domain.
Additionally, integrating multimodal biometric systems and combining fingerprint data with other biometric modalities might present a viable avenue for mitigating the limitations associated with degraded fingerprints.
In this paper, we develop a fingerprint identification approach for low-quality fingerprint images.
The success rate accuracy of the propped algorithm for the low-quality fingerprint images should be significantly better than that of the standard local minutia approach.
The main design of our deep learning approach is based on detecting and extracting the primary correlation during the training and using the correlation feature map to calculate the distance between the low-quality fingerprint images during the predicting phase.
The experimental results show a very promising repulsing and high prediction accuracy.
KeywordsAiFingerprint.
low quality.
deformed fingerprint.
deep learning.
Manuscript received 7 Apr.
revised 19 Jun.
accepted 17 Aug.
Date of publication 30 Nov.
International Journal on Informatics Visualization is licensed under a Creative Commons Attribution-Share Alike 4.
0 International License.
As a result, fingerprints are valuable in forensics and identity verification.
Since the quality of the fingerprint information depends on the quality of the scanned fingerprint picture, accurately and reliably classifying low-quality fingerprints is a challenging data recognition/classification Because recognition success depends on the feature extraction and matching stage, fingerprint recognition systems are susceptible to picture noise and degradation It is often preferred to remove low-quality fingerprint images and replace them with acceptable high-quality ones when using Fingerprint Identification Systems (FIS), which include fingerprint identification approaches because these systems rely on the quality of the fingerprint images.
Acquiring device circumstances .
uch as dirtiness, sensor, and tim.
and individual artifacts .
uch as skin environment, age, skin illness, and pressur.
are among the many elements that impact the fingerprint picture quality.
Partial fingerprint image degradation can occur due to many of these Clear ridges and valleys, and hence the INTRODUCTION Reliable personal information is essential for person identification methods that aim to identify or verify the identity of individuals using various systems.
Participants and users requesting services from the system in question rightfully place a high value on this information in some instances .
The system's main goal is to ensure that only authorized users can access the provided services, including person identification and verification tasks .
The systems are excellent examples: Asynchronous Teller Machines (ATM.
, personal laptops, cellular phones, bank accounts, and protected building access.
However, imposters can easily pull off their schemes without a solid authentication system that includes .
dentification and verificatio.
In biometrics, fingerprints are the most crucial component of human identity.
From birth to death, they remain distinct.
quality dataset.
Kumar.
, et al.
In this study, we provide a hybrid method for low-quality fingerprint identification that combines ensemble learning with local binary patterns (LBP).
LBP features generate local fingerprint patterns from damaged images, and ensemble learning approaches are used for robust classification to increase performance in difficult .
This study offers a privacy-preserving method for matching low-quality fingerprint photos and focuses on privacy issues related to fingerprint identification.
This study delves into cryptographic methods for fingerprint template security, intending to protect user privacy without sacrificing identification accuracy in low-quality data "extractability" of the identification characteristics, are often what determine the quality of fingerprint photographs .
For the identification and verification work, poor-quality fingerprint processing is necessary for improved performance accuracy instead of trying to improve the input picture during pre-processing.
This paper aims to use deep learning approaches to study and enhance low-quality fingerprint recognition and categorization.
Traditional identification methods are sometimes hampered by defective fingerprint pictures, which include noise, distortions, and variances.
this research aims to overcome these issues.
The ultimate goal is to create efficient and trustworthy methods that can accurately use fingerprints to identify and categorize people, even when the data quality is poor.
A method for identifying fingerprints from lowquality photos was suggested in this research.
The propped algorithm outperforms the conventional local minutia technique regarding success rate accuracy when used in lowquality fingerprint photos.
Examining the performance of the proposed system is essential for identifying its behavior when evaluated on low-quality fingerprint photos.
Additionally, we will examine how this behavior is affected by the loss of partial fingerprint information .
Some relevant papers on low-quality fingerprint identification and recognition are summarized here.
According to the authors of .
, he has been working on improving fingerprint identification under challenging environments, such as those with low-quality fingerprints that are distorted and noisy.
The authors suggest an architecture for a deep convolutional neural network (CNN) that can automatically learn discriminative features from fingerprint photos that have been damaged, leading to more accurate and resilient results.
This extensive review examines several fingerprint identification systems, highlighting the difficulties and potential solutions for handling low-quality fingerprint data .
It delves into the shortcomings of conventional approaches, and the improvements ML and DL methods offer to meet these problems.
This study presents an adaptive learning framework developed for low-quality fingerprint recognition.
Improved adaptation to fluctuations in picture quality and resilient performance in demanding settings are achieved by the suggested system, which dynamically adapts its learning parameters based on the quality of input fingerprint data.
their study of uncontrolled settings, .
propose a twopronged fingerprint enhancement and recognition strategy.
The goal of the improvement step is to improve low-quality fingerprint photos, and then we'll use sophisticated machinelearning techniques to identify them.
The research shows that accuracy is improved in difficult circumstances.
In this study, .
investigate techniques for low-quality fingerprint detection that rely on sparse representations.
The suggested solution shows promise when conventional approaches fail by using fingerprint characteristics' sparsity to manage input data distortions and noise.
This study utilizes deep transfer learning algorithms and focuses on the cross-sensor scenario when fingerprint pictures are obtained from several sensors with variable characteristics .
Improved identification performance is achieved by adapting the model to accommodate low-quality fingerprints from various sensors by drawing on knowledge from a high- II.
MATERIALS AND METHOD
Preprocessing Stage: Low-Quality Biometrics Data Preprocessing The first stage of our proposed system is designed to preprocess low-quality biometric data.
Different image processing tools are designed and implemented in this stage, such as image transformation, image binarization, and image The first step of the preprocessing processing stage is image transformation.
In this step, image transformation, normalization, and intensity adjustment are applied to each biometric image .
, .
Fig.
1 Low-Quality Fingerprint Localization Based Low-Quality Biometrics Data During the image transformation, each biometric image is converted to a grayscale image using the Aurgb2gray()Ay function, as is shown in Fig.
Converting low-quality biometric photos to binary images is the second phase of the first stage, which is creation and implementation.
worldwide cutoff is established.
The luminosity of the picture dictates the threshold value.
It will be set to zero if each pixel is less than the threshold.
otherwise, it will remain untouched.
As can be seen in Figure 2 .
, the biometric picture underwent image binarization using a global thresholding approach .
Fingerprint Alignment Step The alignment of fingerprints is the last stage in the fingerprint detection and localization process.
This stage involves isolating each fingerprint object before aligning the extracted binary mask among the x-axes.
This is done after each fingerprint object has been localized using the perfect binary mask for each binary fingerprint.
Once the binary fingerprint mask is aligned correctly, the initial, low-quality fingerprint picture is also adequately aligned .
, .
First, we isolate each fingerprint item using the perfect binary mask for localization.
Then, we align the extracted binary mask among the x-axes.
We begin by extracting the total number of binary objects from the morphological operations using the "regionprops()" function.
In Fig.
, the blue box shows the results of utilizing the "measurements.
BoundingBox" method to extract some features of binary objects, such as the object center and the finger bounding box.
From there, a perfect binary mask "square" is created using "polyshape" and superimposed over the retrieved one, as seen in Figure 2.
Fig.
2 Biometric data pre-processing result.
For SD 302b, baseline data is distributed as PNG files.
For those devices that captured at a resolution other than 196.
pixels per centimeter (PPCM) .
pixels per inch, or PPI), images resampled at 196.
85 PPCM .
PPI) are available.
Samples are from the 302b dataset.
RESULTS AND DISCUSSION
Low-quality Fingerprint Dataset The dataset that is used in our proposed system is the SD302b .
, .
It consists of the operator-assisted rolled fingerprint impressions used as a baseline to quantify Challenger performance and 4-4-2 slap impressions used to ensure the veracity of the reported rolled friction ridge generalized positions.
The images are distributed in the Portable Network Graphics (PNG) image format.
Please refer to NIST Technical Note 2007 .
for complete details.
For details about the N2N Fingerprint Challenge, please refer to NIST Interagency Report 8210 .
The first dataset information is illustrated in the following Table 1.
Evaluation Criteria Our proposed system's low-quality fingerprint detection, localization, and identification evaluation performance is based on two evaluation metrics.
The proposed system produces different versions from the same dataset.
For instance, the first approach.
Aulow-quality fingerprint detection and localization approach,Ay produces three datasets.
The first one is the original SD 302b after detection and localized each fingerprint, while the second dataset is the aligned SD 302b dataset, and the third one is the localized SD 302b.
Each dataset is used as a training/testing dataset for the third approach.
Aulow-quality fingerprint identification and classification approachAy.
For this task, different criteria such as Accuracy, precision.
Recall, and F1-measure and are used to evaluate our first approach
TABLE I
LOW-QUALITY FINGERPRINT DATASET DESCRIPTION (SD-302B)
Dataset Name Sd302b Device Resolution Image No.
As shown in Fig.
4, each of the three biometric images from the SD-302b produces 10 individual low-quality fingerprint These are collected and isolated automatically through our first approach in one individual class, for a total of 92 classes.
Each class has ten individual low-quality fingerprint images .
, .
Low Quality Fingerprint Detection and Isolation Experimental Results Our approach's experimental result is based on using all biometric images from the SD-302b dataset.
In this stage, our proposed method uses two biometric images from the SD302b dataset to detect and localize each class .
fingerprint image.
The dataset is pre-processed by splitting it into multiple classes based on the dataset configuration.
Samples from the dataset are illustrated in Table 2.
Low Quality Fingerprint Aligning and Localization Experimental Results The second experimental result of our first approach is based on producing different versions of the first dataset that are automatically generated through our first approach.
Each fingerprint automatically detected in this stage is aligned and localized .
, .
TABLE II
LOW QUALITY FINGERPRINT DETECTION AND PREPARATION EXPERIMENTAL
RESULTS USING SD-302B LOW-QUALITY FINGERPRINT DATASET
No.
Classes No.
of Images per Class Image Total No.
Class 1 Class 2 Fig.
4 Sample images from the detection and preparation of experimental results using the SD-302b Low-quality fingerprint dataset.
Table 3 illustrates the total number of images for each dataset, and some samples from different datasets are also TABLE i LOW-QUALITY FINGERPRINT DETECTION.
ALIGNING.
AND LOCALIZATION
EXPERIMENTAL RESULTS USING SD-302B LOW-QUALITY FINGERPRINT
DATASET.
Generated Dataset Original Aligned Localized Total No.
of Classes Images Per Class Image Total No.
of images Low-quality Fingerprint Image Aligning Experiential Results The first part of this experimental result is the low-quality fingerprint image alignment.
In this part, each individual lowquality fingerprint image is automatically aligned using our first approach.
Aulow-quality fingerprint image aligning and localization approach.
Ay Samples from the aligned low-quality fingerprint dataset are shown in Fig.
This generates three different datasets .
riginal, aligned, and localize.
with a total of 2760 low-quality fingerprint Fig.
5 Sample images from low-quality fingerprint aligning experimental results using SD-302b Low-quality fingerprint dataset.
has been used to identify and classify low-quality The main task of this approach is to find the main correlation between the low-quality fingerprints through the training phase.
Then, the learning parameters .
earned correlation ma.
are used later to predict the correct label for the low-quality fingerprint in the testing dataset.
Each dataset .
riginal, aligned, and localize.
is split into training and testing datasets using 70% training and 30% Table 4 below describes the training testing scheme for the third proposed approach.
Low-quality Fingerprint Image Localization Experiential Results The second part of this experimental result is the lowquality fingerprint image localization.
In this part, each lowquality fingerprint image is automatically localized, and the ROI of each fingerprint is automatically extracted and localized using our first approach.
Aulow-quality fingerprint image aligning and localization approachAy.
Samples from localized low-quality fingerprint datasets.
The proposed deep learning approach implemented in our third approach
for low-quality fingerprint identification and classification TABLE IV
SHOW THE DATA PARTITIONS OF OUR PROPOSED SYSTEM USING SD-302B
LOW-QUALITY FINGERPRINT DATASET
Dataset
Data Partition Training and Validation 70% Testing 30% Total
Localized low-quality fingerprint dataset
Number of images QUALITY FINGERPRINT IDENTIFICATION AND CLASSIFICATION APPROACH
USING THE SD-302B DATASET
TrainingValidation 70% Testing 30% Total TrainingValidation 70% Testing 30% Original low-quality fingerprints dataset Aligned low-quality fingerprint dataset Number of Experimental Results using the Original SD-302b Lowquality Fingerprint Dataset The first dataset used is the original detected low-quality fingerprint dataset.
The training and testing scheme used in this experimental result, such as the training parameters and number of epochs, learning rate, and the accuracy of the first experiment, are explained in Table 6.
To come up with the best training/testing parameters, we tried different learning methods .
The best learning parameters with the highest accuracy were using 400 epochs, with 64 learning patches and 4400 iterations per epoch.
For this reason, we used these parameters as our final learning platform for other experimental results.
TABLE V
THE WHOLE DATASET THAT HAS BEEN USED IN THE PROPOSED LOW-
Data Partition Total TrainingValidation 70% Testing 20% Total Total This approach trains our proposed deep learning model on three different datasets, as shown in Table 5 below.
Dataset
Data Partition Number of TABLE VI
TRAINING SCHEME AND PARAMETERS FOR THE FIRST APPROACH USING THE ORIGINAL SD-302B LOW-QUALITY FINGERPRINT DATASET
No.
of Epochs Batch Size Training Time No.
of Iteration Iteration Epoch Learning Rate Accuracy 15 min 31 min 18 min 15 min 13 min 71 min 33 min 24 min As discussed above, the training loss and the validation accuracy for the low-quality fingerprint identification and classification were based on using 200 epochs only, with 128 learning patch and 200 iterations per epoch.
The confusion metrics of the testing dataset is also shown in the following Fig.
Some testing samples of the low-quality fingerprint identification and classification based on using 200 epochs only, with 128 learning patch and 200 iterations per epochs is shown below in the following Fig.
Fig.
Training loss and validation accuracy for the low-quality fingerprint identification and classification approach using SD-302b Low-quality fingerprint dataset using 200 epochs, 64 learning patches and 1100 iterations, .
Confusion metrics for the testing dataset for the low-quality fingerprint identification and classification approach using SD-302b Low-quality fingerprint dataset using 200 epochs, 64 learning patches and 1100 iterations Fig.
7 Testing samples of the low-quality fingerprint identification and classification approach using SD-302b Low-quality fingerprint dataset using 200 epochs, 128 learning patches, and 200 iterations.
shown in the following Fig.
Some testing samples of the low-quality fingerprint identification and classification based on using aligned SD-302b Low-quality fingerprint dataset is shown below in the following Fig.
Experimental Results using the Aligned SD-302b Lowquality Fingerprint Dataset The second dataset used is the aligned low-quality fingerprint dataset.
The training and testing scheme used in this experimental result is Table 7, which explains the training parameters, number of epochs, learning rate, and accuracy of the first experimental experiment.
As discussed above, the training loss and the validation accuracy of using 400 epochs only, with 64 learning patches and 4800 iterations per epoch are shown below in the following Fig.
The confusion metrics of the testing dataset are also TABLE VII
TRAINING SCHEME AND PARAMETERS FOR THE FIRST APPROACH USING
THE ALIGNED SD-302B LOW-QUALITY FINGERPRINT DATASET.
No.
Epoch Batc Size Trainin g Time No.
Iteratio Iteratio Epoch Learnin g Rate Accurac 38 min Fig.
Training loss and validation accuracy for the low-quality fingerprint identification and classification approach using aligned SD-302b Low-quality fingerprint dataset with 400 epochs, 64 learning patches and 4800 iterations, .
Confusion metrics for the testing dataset for the low-quality fingerprint identification and classification approach using SD-302b Low-quality fingerprint dataset using 50 epochs, 16 learning patches and 2000 iterations per epoch.
Fig.
9 Testing samples of the low-quality fingerprint identification and classification approach using aligned SD-302b Low-quality fingerprint dataset using 400 epochs, 64 learning patches, and 4800 iterations.
The third dataset used is the aligned low-quality fingerprint Table 8 explains the training and testing scheme used in this experimental result, including the training parameters.
Experimental Results using the Localized SD-302b Lowquality Fingerprint Dataset number of epochs, learning rate, and accuracy of the first As discussed above, the training loss and the validation accuracy of using 400 epochs only, with 64 learning patches and 4800 iterations per epoch is shown below in the following Fig.
The confusion metrics of the testing dataset are also shown in the following Fig.
Some testing samples of the low-quality fingerprint identification and classification based on using aligned SD-302b Low-quality fingerprint dataset is shown below in the following Fig.
TABLE Vi
TRAINING SCHEME AND PARAMETERS FOR THE FIRST APPROACH USING THE
LOCALIZED SD-302B LOW-QUALITY FINGERPRINT DATASET.
No.
Epoch Batc Size Trainin g Time No.
Iteratio Iteratio n Epoch Learnin g Rate Accurac 37 min Fig.
Training loss and validation accuracy for the low-quality fingerprint identification and classification approach using localized SD-302b Low-quality fingerprint dataset with 400 epochs, 64 learning patches and 4800 iterations, .
Confusion metrics for the testing dataset for the low-quality fingerprint identification and classification approach using localized SD-302b Low-quality fingerprint dataset using 50 epochs, 16 learning patches and 2000 iterations per Fig.
11 Testing samples of the low-quality fingerprint identification and classification approach using aligned SD-302b Low-quality fingerprint dataset using 400 epochs, 64 learning patches and 4800 iterations.
fingerprint, such as the original detected one, aligned one, and localized one.
IV.
CONCLUSION
In conclusion, pursuing effective low-quality fingerprint identification and classification using machine learning and deep learning approaches has been challenging yet critical.
This thesis has delved into the issue of low-quality fingerprint data and its impact on the accuracy of identification and classification systems.
The findings and insights gained throughout this research shed light on the complexities and limitations of leveraging these advanced techniques in scenarios where fingerprint quality is compromised.
Exploring traditional machine learning methods revealed that their performance is significantly hindered when confronted with low-quality fingerprint images.
The sensitivity to noise, distortions, and variations in such data proved a major obstacle in achieving reliable identification and classification Fig.
12 Overall testing performance of the low-quality fingerprint identification and classification approach using original, aligned, and localized SD-302b Low-quality fingerprint dataset.
In general.
Fig.
12 illustrates the overall performance of the low-quality fingerprint identification and classification approach using different datasets of the SD-302b low-quality outcomes.
Despite incorporating feature engineering and extraction techniques, these methods exhibited limitations in handling the intricacies of degraded fingerprint images.
Integrating deep learning approaches, particularly convolutional neural networks (CNN.
, showcased promise in addressing some of the challenges posed by low-quality CNN demonstrated an inherent ability to learn hierarchical representations directly from raw input data, enabling them to capture complex patterns and features even in noise and distortions.
However, the effectiveness of deep learning models is contingent upon the availability of large and diverse training datasets, which may be scarce for lowquality fingerprint scenarios .
REFERENCES