2024 Second International Conference on Networks, Multimedia and Information Technology (NMITCON)

Recognition of Fingerprint Images using CNN for

2024 Second International Conference on Networks, Multimedia and Information Technology (NMITCON) | 979-8-3503-7289-2/24/$31.00 ©2024 IEEE | DOI: 10.1109/NMITCON62075.2024.10698773

Cybercrime Detection System

1st T.Kavitha 2nd Bhimaraya Patil 3rd S. Saraswathi
Dept. of Computer Engineering Dept. of Computer Engineering Dept. of Computer Science Engineering Sri
New Horizon College of engineering New Horizon College of Engineering Sivasubramaniya Nadar College of Engineering,
Bangalore, India Bangalore, India Chennai, India.
[email protected] [email protected] [email protected]
4th S. Rajarajeswari 5th Anita Patil
School of computer science and engineering Dept., of Information Science
Vellore Institute of Technology City Engineering College,
Vellore, India. Bengaluru, India.
[email protected] [email protected]

Abstract- Fingerprint identification plays a crucial role in crime agencies to swiftly identify suspects and solve cases. [4][5]
detection, but manual analysis is time-consuming and error- This limitation has motivated researchers to explore
prone. This research aims to develop an automated system for automated techniques leveraging machine learning
accurate and efficient fingerprint matching using machine algorithms, aiming to enhance the efficiency and reliability of
learning. The proposed approach involves image acquisition, fingerprint recognition systems. [6] The primary objective of
preprocessing, feature extraction, and classification. Fingerprint
this work is to develop a robust fingerprint image
images are enhanced, binarized, and segmented. Minutiae (ridge
endings and bifurcations) are extracted as features. A identification system for crime detection using advanced
convolutional neural network (CNN) is trained on a dataset of machine learning techniques, specifically convolutional
fingerprint images for classification. By comparing algorithms, neural networks (CNNs). [7] By automating the process of
Random Forest achieved over 81% accuracy, outperforming feature extraction, classification, and matching, the proposed
Decision Trees (nearly 75% accuracy). The trained CNN model system aims to provide law enforcement agencies with a
can accurately match fingerprints from crime scenes with reliable and efficient tool for identifying suspects based on
potential suspects, providing reliable evidence. The developed fingerprint evidence from crime scenes. [8] The key
system offers law enforcement agencies an automated and contributions of this research include an in-depth analysis of
accurate fingerprint identification tool, improving crime-solving
various CNN architectures and their performance in
efficiency and reducing manual effort. The proposed approach
showcases the potential of machine learning in forensic fingerprint classification, as well as the development of a user-
applications. friendly interface for seamless interaction with the fingerprint
identification system. [9][10][11]. The paper isorganized as
Keywords- Random Forest Algorithm, Ensemble Learning, Image follows: Section II discusses about the literature survey and
Analysis, Pattern Recognition, Fingerprint Matching. Section III explains about the proposed work. Section IV gives
the overview of implementation. Section V discusses about the
I. INTRODUCTION results and discussion and finally conclusion in given in
Fingerprint recognition has long been considered the gold section VI.
standard for personal identification in forensics, having II. LITERATURE SURVEY
provided crucial evidence in numerous criminal investigations Fingerprint recognition and identification have been
over the past century. [1] Despite the emergence of DNA extensively researched for crime detection and forensic
fingerprinting, fingerprint analysis remains universally used applications, employing diverse techniques and approaches.
due to the unique and durable nature of friction ridge patterns Several studies have focused on image enhancement, feature
found on the skin. These intricate formations, developed extraction, and classification methods. Murthy et al. [1]
during the third to fourth month of fetal development, serve to proposed a methodology involving image enhancement,
enhance grip and prevent slippage when handling objects. thresholding, skeletonization, and minutiae extraction using
[2][3]. The distinctive arrangement and combination of ridge the Cross-Number method, along with the AlexNet CNN
characteristics make fingerprints an ideal biometric marker for architecture. Yilmaz [2] utilized Generative Adversarial
establishing human identity. However, the process of Networks (GANs) and one-shot learning for feature extraction
fingerprint identification for crime detection is not without its and differentiation. Militello et al. [3] compared pre-trained
challenges. Crime scene fingerprints are often partial, CNN architectures like AlexNet, GoogLeNet, and ResNet for
distorted, or obscured by noise, making them difficult to fingerprint classification, while Wen et al. [4] proposed a
analyze and match accurately. Traditional manual analysis lightweight CNN structure based on singularity Region of
methods are labor-intensive, time-consuming, and prone to Interest (ROI).
human error, hindering the ability of law enforcement

979-8-3503-7289-2/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.
Other studies have explored texture features and classifier [12] analyzed fingerprint recognition using the Cross-Number
algorithms for fingerprint classification. Hameed et al. [5] method for minutiae extraction and counting[14-15]. Fig 1 and
presented a method based on Histogram of Oriented Gradient Fig 2 shows the comparison of various algorithem with respect
(HOG) and Segmentation-based Feature Texture Analysis to different performance metrics like memory usage, training
(SFTA) features, combined with Discriminant Analysis and time per epoch, accuracy and loss. From the graph it is
Gaussian Discriminant Analysis classifiers. Li et al. [9] observed that CNN given more accuracy with less loss and at
proposed a fingerprint liveness detection method using fine- the same time it needs less memory. So for the proposed
grained feature fusion of SIFT, HOG, and LBP features, with system CNN is used to improve efficiency.
an SVM classifier. Uliyan et al. [7] introduced an anti-
spoofing method using a Deep Boltzmann Machine (DBM) III. PROPOSED WORK
and the KNN classifier to combat spoof forgeries. Techniques
for automated latent fingerprint identification have also been
A. SYSTEM ARCHITECTURE
investigated. Singla et al. [8] reviewed existing techniques and
The system architecture of the proposed system is shown in
highlighted the need for a completely automated system,
Fig 3. It encompasses several distinct phases, including image
addressing challenges in segmentation, overlapped fingerprint
acquisition, pre-processing, feature extraction, and
separation, and enhancement. Win et al. [6] surveyed
classification. The initial step involves the collection of
fingerprint classification algorithms and their applications in
fingerprint images utilizing specialized equipment.
criminal investigations, discussing methods like K-Means
Subsequently, these images undergo a pre-processing stage,
Algorithm, Apriori Algorithm, minutiae extraction, thinning,
during which any noise or artifacts that could potentially
and vein recognition[10-11].
interfere with the analysis are removed. The pre-processed
images are then subjected to sophisticated feature extraction
techniques, which identify the unique characteristics of the
fingerprints, such as ridge endings, bifurcations, and minutiae
points. Once the relevant features have been extracted, they
are utilized to train a machine learning algorithm, such as a
convolutional neural network (CNN), to classify the
fingerprint images. The algorithm leverages these features to
create a model that can distinguish between matching and non-
matching fingerprints with a high degree of accuracy

Fig 1.Training Time and Memory Usage of Various Model

Fig 3. Block Diagram

1. Image Enhancement:
Image segmentation is performed by partitioning the grayscale
fingerprint image based on pixel characteristics, separating the
regions of interest. Advanced filtering techniques and contrast
enhancement methods are applied for noise removal and ridge-
valley visibility improvement, respectively. Additionally,
Fig 2.Accuracy and Loss of Various Model
Random Forest models trained on diverse fingerprint datasets
Several studies have proposed unique approaches to are employed for tailored image enhancement.
fingerprint identification and matching. Amer et al. [10] The Random Forest algorithm constructs multiple decision
presented a model fusing morphological and neural network trees, where each tree is represented by:
approaches for latent fingerprint segmentation, patch f(x) = Σ w_i * I(x ∈ R_i) (1)
classification, and reconstruction. Witefee and Al-Asad [12- Here, x is the input feature vector, R_i represents the partitions
14] introduced an algorithm based on the angle direction of of the feature space, I(x ∈ R_i) is an indicator function that
minutiae objects, utilizing the Scale Invariant Feature takes the value 1 if x falls within the region R_i, and 0
Transformation (SIFT) method for matching. Awasthi et al. otherwise, and w_i is the prediction value associated with the

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.
region R_i. The final prediction is the aggregation of multiple 5. Matching: The Random Forest algorithm is an ensemble
decision trees' predictions. learning method that combines multiple decision trees to
2. Cutting-Edge Segmentation Strategies: Image binarization improve accuracy and robustness. The mathematical equation
and segmentation are crucial steps in fingerprint analysis. for the Random Forest model can be represented as:
Binarization converts the grayscale image into a binary f(x) = 1/M * Σ(m=1 to M) fm(x) (4)
format, representing ridges as black and valleys as white. One Where f(x) is the final prediction of the Random Forest model
common method is Otsu's algorithm, which determines an for input x,
optimal threshold value, T, by minimizing the intra-class M is the number of decision trees in the ensemble
variance: fm(x) is the prediction of the m-th decision tree for input x
σ_w^2(T) = ω_1(T)σ_1^2(T) + ω_2(T)σ_2^2(T) (2)
Here, ω_1(T) and ω_2(T) are the class weights, and σ_1^2(T)
and σ_2^2(T) are the variances of the two classes separated by
the threshold T.
Segmentation separates an image containing multiple
fingerprints into individual images, each containing a single
fingerprint. If the initial image has N distinct fingerprints, the
segmentation process extracts N separate images. This can be Fig 4. Image Enhancement
achieved using techniques like connected component analysis
or contour detection, which identify and separate connected B.MODULE DESCRIPTION
regions or contours in the binary image. 1. Image Enhancement: The input fingerprint image
3. Feature Extraction: Feature extraction is a crucial step in undergoes segmentation to separate the fingerprint area
fingerprint analysis, where distinctive minutiae details are from the background. Next, normalization is applied to
identified from the thinned fingerprint image. The most enhance the image quality. Fig. 4: Orientation
prominent and stable minutiae types used are bifurcations and estimation is then performed to determine the ridge
ridge endings. The Cross-Number (CN) concept is a widely orientation in the fingerprint. Ridge frequency
employed method for minutiae extraction, which involves estimation is carried out to calculate the frequency of
scanning the local neighborhood of each ridge pixel in the the ridges. These steps are crucial for accurate
skeleton image using a 3×3 window. The CN for a ridge pixel fingerprint feature extraction and matching in
is calculated as: biometric recognition systems.
CN = 0.5 * Σ |P_i - P_{i+1}| (3) 2. Preprocessing: The input image undergoes gray
Where P_i and P_{i+1} are the values of two adjacent pixels conversion to obtain a grayscale representation. Next,
in the 3×3 window. The skeleton image is then classified into filtering operations are applied to remove noise and
bifurcations (CN = 3) and ridge endings (CN = 1) based on the enhance the image quality. Fig. 5: Binarization is then
CN properties, extracting the minutiae details. performed to convert the grayscale image into a binary
format, where pixels are represented as either black or
4. Training and testing Model white. These preprocessing steps of gray conversion,
After minutiae extraction, the fingerprint data is used for filtering, and binarization are crucial for effective
training a CNN (convolutional neural network). If the dataset fingerprint image analysis and feature extraction in
has 2000 fingerprint images, 1500 are used for training and the biometric recognition systems.
remaining 500 for testing. An investigator takes fingerprints 3. Minutiae Extraction: The input image file undergoes
from the crime scene and compares them to a database of generative image maps creation. Binarization is then
previous criminals. After preprocessing the images, the CNN performed to convert the image into a binary format.
extracts features and ranks the identified criminals based on Fig. 6: Minutiae (ridge endings and bifurcations) are
similarity to the crime scene fingerprints, providing an detected from the binarized image. False minutiae are
accuracy score. The CNN's deep learning ability allows it to removed based on certain criteria, and neighboring
learn relevant features from the images, outperforming SVM ridges are counted to assess the quality of the remaining
(support vector machine) classifiers. Using the CNN classifier, minutiae. Finally, an output minutiae file containing
an improved fingerprint identification accuracy of 99% is the extracted genuine minutiae is generated for further
achieved. Here is the 6-line paragraph condensed into 3 lines: processing in fingerprint recognition systems.
For testing the fingerprint identification model, a separate test
set of 500 fingerprint images distinct from the 1500 training
images is used. After preprocessing, the test fingerprints are
passed through the trained CNN model which extracts
features, matches against the database, and outputs ranked
matches with similarity scores. The top matches are compared
to ground truth labels to evaluate the CNN model's accuracy
on the test set using metrics like precision, recall, and F1-
score. Fig 5. Preprocessing

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.
Fig 6. Minutiae Extracgtion
Fig 8.Fingerprint Data Flow Diagram
4. Matching:
Fig. 7 depicts the fingerprint recognition process. The input C. .USE CASE
fingerprint undergoes acquisition, preprocessing, and feature Fig. 9 depicts a Criminal Detection System with a use case
extraction stages. A convolutional neural network (CNN) is diagram illustrating the interactions between users and the
trained on a dataset, with one subset used for training and system. Users can load fingerprint images into the system,
another for testing. During an investigation, crime scene which can then be added to the database. Additionally, users
fingerprints are preprocessed, and the trained CNN extracts can match an input fingerprint against the database to identify
features. These features are matched against a database of potential matches or no matches. The system facilitates
known offenders' fingerprints through a dedicated matching fingerprint management and identification, aiding criminal
module. Potential matches are ranked based on similarity, and investigations through fingerprint evidence. For example, if
an accuracy metric quantifies the confidence level of each there are 100 instances of the Tented Arch pattern in the
match, leading to the final decision. testing set, and the Random Forest classifier correctly
identifies 91 of them, the accuracy for the Tented Arch pattern
would be 0.91 or 91%.

IV. SYSTEM IMPLEMENTATION

The proposed system is implements in a 4GB GPUwith

Python intergrated development environment Version 3.7.
Major libraries used are TensorFlow, Numpy, Pandas and
Open CV.
Fig 7. Matching Process A. Acquisition: Fingerprints (Fig10) are unintentionally left
behind as abrasion impressions on various surfaces. The
B. DATA FLOW AND DATA SET images of these fingerprints can be obtained through
In a fingerprint recognition system(Fig 8), the data flow begins various methods, ranging from chemical processing
with an input fingerprint image that undergoes acquisition and techniques to precision photography. The images
preprocessing for quality enhancement. Distinctive features acquired directly from crime scenes, known as crime
are then extracted from the preprocessed image and matched scene images (CSI), play a crucial role and serve as
against a template fingerprint through dedicated algorithms. evidence in criminal cases. However, the fingerprint
Based on the similarity score, a decision is made regarding the information contained in these crime scene images may
identity or authentication status of the input fingerprint, be partial or challenging to identify, resulting in low-
enabling accurate fingerprint recognition. quality fingerprint images that may have limited value for
For training and evaluating the fingerprint identification purposes.
identification model, Proposed system utilized the publicly
available FVC2004 database from the Biometric Systems Lab
at University of Bologna. This dataset contains a total of 8000
fingerprint images acquired using an optical sensor, equally
divided into four distinct databases (DB1-DB4). Each
database consists of 800 fingerprint images from 100 different
fingers, with 8 impressions per finger. I split this dataset into
a training set of 1500 images and a test set of 500 images while
ensuring subject independence between the two sets. Prior to
feeding into the model, the fingerprint images underwent
preprocessing steps like normalization, orientation estimation,
ridge frequency estimation, and Gabor filtering to enhance the
ridge structures and extract robust features.

Fig 9.Criminal Detection System

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.
 INPUT IMAGE GRAY IMAGE PROCESSED IMAGE
Fig 10.Acquisition

B.Preprocessing: To enhance classification accuracy, data pre-

processing (Fig 11) steps are crucial. Fingerprint images are
resized to match CNN input dimensions and converted from
grayscale to RGB format. Singularity ROI extraction, the
initial step, isolates the pattern distribution around singularity
points for classification input, reducing computational Fig 13.Minutia mapping
complexity while maintaining performance. The original
512x512 pixel images have 256 grayscale levels. D.Matching and Verification:Subsequent to minutiae extraction
(Fig 13), fingerprint data undergoes CNN training, utilizing
1500 images for training and 500 for testing from a 2000-
image dataset.After pre-processing, the CNN system extracts
features, ranking identified individuals by similarity to crime
scene fingerprints and providing an accuracy score. In the
verification stage, the system confirms match accuracy by
ensuring the similarity score exceeds a predetermined
threshold. If exceeded, it identifies the individual and provides
relevant criminal investigation information.The verification
step is crucial, validating matching accuracy and determining
identity by comparing the similarity score against a threshold
to ascertain fingerprint match or mismatch.
Fig 11.Preprocessing V. RESULT AND DISCUSSION
C.Feature Extraction: Minutia extraction, or feature extraction
The project aims to create an automated machine learning-
(Fig 12), identifies distinctive fingerprint features from
based system for fingerprint identification in crime detection,
thinned images, known as minutiae. It is challenging to
accurately represent crime scene fingerprint input.A addressing the inefficiencies and inaccuracies of manual
analysis. It involves collecting fingerprint images, pre-
fingerprint's valid representation lies in its minutiae detail
processing them to remove noise, and extracting unique
pattern, with around 150 local characteristics identified as
features such as ridge endings and bifurcations. A
minutiae details.Only bifurcations (ridge splits into two paths)
convolutional neural network (CNN) is used for classification,
and ridge endings (ridge curve termination points) are
and Random Forest algorithms enhances accuracy (Fig 14 &
commonly used due to stability and robustness.The Cross-
Number (CN) concept, involving scanning ridge pixels with a Fig 15). Image enhancement techniques, including filtering
and contrast improvement, improves fingerprint quality. The
3x3 window, is the most widely used minutiae extraction
ultimate goal is to provide law enforcement with a reliable tool
method.Extracted minutiae are classified as bifurcations or
for solving crimes efficiently.
ridge endings based on their CN properties.

Fig 12.Minutia Extarction Fig 14.Accuracy ploting of machine learning

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.
 [9] S. S. Hameed, I. T. Ahmed and O. M. A. Okashi, "Real and altered
fingerprint classification based on various features and classifiers,"
Accuracy Value Computers, Materials & Continua, vol. 74, no.1, pp. 327–340, 2023.
[10] W. Jian, Y. Zhou and H. Liu, "Lightweight Convolutional Neural
Network Based on Singularity ROI for Fingerprint Classification," in
Pattern Name

Central Pocket Loop Whorl

IEEE Access, vol. 8, pp. 54554-54563, 2020.
Plain Whorl [11] M. Liu and P. Qian, "Automatic Segmentation and Enhancement of
Radial Loop Latent Fingerprints Using Deep Nested UNets," in IEEE Transactions
on Information Forensics and Security, vol. 16, pp. 1709-1719, 2021,
Tented Arch doi: 10.1109/TIFS.2020.3039058.
0 0.2 0.4 0.6 0.8 1 [12] G. Dency Flora, G. Sekar, R. Nivetha, R. Thirukkumaran, D.
Silambarasan and V. Jeevanantham, "An Optimized Neural Network for
Accuracy Value Content Based Image Retrieval in Medical Applications," 2022 8th
International Conference on Advanced Computing and Communication
Fig 15.Classification Fingerprint pattern And Accuracy With Random Forest Systems (ICACCS), Coimbatore, India, 2022, pp. 1560-1563, doi:
10.1109/ICACCS54159.2022.9785151.
[13] Kavitha, T., Malini & Senbhagavalli. "Deep Learning for Emotion
VI. CONCLUSION AND FUTURE ENHANCEMENT Recognition." Handbook of Research on Computer Vision and Image
Processing in the Deep Learning Era, edited by A. Srinivasan, IGI
The Propossed work differs from other research papers by a Global, 2023, pp. 56-91.
fusion of morphological and neural network approaches for [14] S Pavithra, N K Shreyashwini, H S Bhavana, G Nikhitha, T Kavitha,
latent fingerprint segmentation and identification. The “Hand-Drawn Electronic Component Recognition Using ORB”,
Procedia Computer Science, Volume 218, 2023, Pages 504-513, ISSN
morphological method finds the fingerprint region, which is 1877-0509.
divided into patches, and a CNN classifies the patches as [15] Priya, Alikapati Keerthi and N, Gaganashree and K.N, Hemalatha and
fingerprint or non-fingerprint. This approach achieved an Chembeti, Janaki Sutha and T, Kavitha, “AI Based Online Hand Drawn
improved fingerprint identification accuracy of 80% using the Engineering Symbol Classification and Recognition”, 9th Intl. Conf. On
CNN classifier. For future enhancement, explore advanced Innovations In Electronics & Communication Engineering(ICIECE -
2021) 13-14 AUGUST, 2021.
fingerprint sensors and high-resolution scanners to capture
[16] Kavitha, T., Sandhya, M.K., Subashini, V.J., & Srikanth, P. (Eds.).
more detailed images. Combining proposed approach with (2024). Secure Communication in Internet of Things: Emerging
deep learning algorithms like CNNs trained on larger datasets Technologies, Challenges, and Mitigation (1st ed.). CRC Press.
could further improve accuracy and reduce false
positives/negatives. Additionally, integrating machine
learning algorithms for feature extraction and classification
could lead to faster and more robust fingerprint identification.

REFERENCES

[1] Baskar M, Rajagopal RD, BVVS P, Babu JC, Bartáková GP, Arulananth
TS. Multi-region minutiae depth value-based efficient forged finger
print analysis. Plos one. 2023 Nov 16;18(11).
[2] Nogueira, Rodrigo Frassetto, Roberto de Alencar Lotufo, and Rubens
Campos Machado. "Fingerprint liveness detection using convolutional
neural networks." IEEE transactions on information forensics and
security 11.6 (2016): 1206-1213.
[3] Kaur, Manvjeet, et al. "Fingerprint verification system using minutiae
extraction technique." International Journal of Computer and
Information Engineering 2.10 (2008): 3405-3410.
[4] Deepthi Murthy T S, Chirakala Sravya , Chadavala Sai Niharika,
Dudekula Sai Kumar , Dwarsala Ramanjaneyula Reddy: “Fingerprint
Image Recognition for Crime Detection”,Turkish Journal of Computer
and Mathematics Education Vol.12 No.12 (2021), 2230-2237.
[5] Kumar, Tarun, and Ravi Shanker Garg. "The recognition of latent
fingerprints using swarm intelligence based hybrid approach." Int J
Emerg Technol 11.5 (2020): 90-97.
[6] Pavithra, R., and K. V. Suresh. "Fingerprint image identification for
crime detection." 2019 International Conference on Communication and
Signal Processing (ICCSP). IEEE, 2019.
[7] Khin Nandar Win, Kenli Li, Jianguo Chen, Philippe Fournier Viger,
Keqin Li,Fingerprint classification and identification algorithms for
criminal investigation: A survey,Future Generation Computer Systems,
Volume 110, 2020, Pages 758-771, ISSN 0167-739X.
[8] Ibrahim Yilmaz Department of Computer Science Tennessee
Technological University “Enhanced Fingerprint Identification
Approach Using GAN and One-shot learning Techniques”
arXiv:2208.05615v1 [cs.CV] 11 Aug 2022.

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on April 07,2025 at 07:27:11 UTC from IEEE Xplore. Restrictions apply.