Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Criminal Face Recognition Using GAN

Anitta George Krishnendu K A Anusree K
Dept. of Computer Science Dept. of Computer Science Assistant prof.
Sahrdaya College of Engineering and Sahrdaya College of Engineering and (Dept. of Computer Science)
Technology Thrissur, India Technology Thrissur, India Sahrdaya College of Engineering and
Technology Thrissur, India

Adira Suresh Nair Hari Shree

Dept. of Computer Science Dept. of Computer Science
Sahrdaya College of Engineering and Technology Sahrdaya College of Engineering and Technology
Thrissur, India Thrissur, India

Abstract:- Forensics and security at present often use distribution. The generative model can be thought of as
low technological resources. Security measures often fail analogous to a team of counterfeiters, trying to produce
to update with the upcoming technology. This project is fake currency and use it without detection, while the
based on implementing an automatic face recognition of discriminative model is analogous to the police, trying to
criminals or specific targets using machine-learning detect the counterfeit currency. Competition in this game
approach. Given a set of features to a Generative drives both teams to improve their methods until the
Adversarial Network(GAN), the algorithm generates an counterfeits are indistinguishable from the genuine articles.
image of the target with the specified feature set. The This framework can yield specific training algorithms for
input to the machine can either be a given set of features many kinds of model and optimization algorithms. The
or a set of portraits varying from frontals to side special case when the generative model generates samples
profiles from which these features can be extracted. The by passing random noise through a multilayer perceptron,
accuracy of the system is directly proportional to the and the discriminative model is also a multilayer perceptron
number of epochs trained in the network. The is explored. This special case is referred to as adversarial
generated output image can vary from primitive, low nets. In this case, training is performed on both models
resolution images to high quality images where features using only the highly successful backpropagation and
are more recognizable. This is then compared with a dropout algorithms and samples from the generative model
predefined database of existing people. Thus, the target using only forward propagation. No approximate inference
can immediately be recognized with the generation of or Markov chains are necessary.
an artificial image with the given biometric feature set,
which will be again compared by a discriminator II. DRAWBACKS OF EXISTING SYSTEM
network to check the true identity of the target.
 Human-made sketches may be ambiguous and unclear.
I. INTRODUCTION  The primitive sketches of a target may be black and
white or monochrome. Features which are colour
The promise of deep learning is to discover rich, dependent will not be significant.
hierarchical models that represent probability distributions  Some features can be changed by culprit/target. E.g.
over the kinds of data encountered in artificial intelligence hair colour, hairstyle, etc.
applications, such as natural images, audio waveforms  Face recognition by humans takes more time if the
containing speech, and symbols in natural language target face has to find a match without a database.
corpora. So far, the most striking successes in deep learning
have involved discriminative models, usually those that III. METHOD
map a high-dimensional, rich sensory input to a class label.
These striking successes have primarily been based on the The proposed method presents a detailed deep
backpropagation and dropout algorithms, using piecewise learning-based model for face identification and detection.
linear units which have a particularly well-behaved We are proposing this method specifically for criminal
gradient. Deep generative models have had less of an identification. An image of the person, who you want to
impact, due to the difficulty of approximating many identify is given as the input to the machine or generative
intractable probabilistic computations that arise in network. This image may be uploaded by the investigating
maximum likelihood estimation and related strategies, and crew through an android application which can be installed
due to difficulty of leveraging the benefits of piecewise into their mobile phones.
linear units in the generative context. A new generative
model estimation procedure is proposed that side-steps Our goal is to determine the target’s identity by
these difficulties. In the proposed adversarial nets utilizing the visual clues obtained via the camera. The input
framework, the generative model is pitted against an image is fed into the Generative Adversarial Network
adversary: a discriminative model that learns to determine (GAN) via android app. The image is taken into the
whether a sample is from the model distribution or the data generator first after preprocessing, the image features that

IJISRT20JUN1116 www.ijisrt.com 1526

Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
are visible are stored as a feature set f(x) for the input
image x. The features set is used to generate labels of the
same target with slightly different features. This would
describe the target that they changed their appearance at
present and made variations such as change of hair colour,
eye colour, etc. Generative Adversarial Networks are
powerful enough to identify the target even in their best
disguise since their facial features remain the same.

Fig 1:- Working of a GAN network

The generator module of the GAN network takes the

features of an image as an input and produces a set of labels
for each face found. This output image is fed into the
discriminator module of the same network. It is the
responsibility of the discriminator to take the generator’s
output and check if the features of the given target match
with any of the ones in the database. If there is a match, the
image will be invalidated and the identity of the target will
be returned. Thus the network will extract as many visible
features possible from the images x= 1 to n, and combine
all features to produce the superset of features f(x). The
identity of the target together with their original image is
returned as the output through android app. The generator
has two other submodules called encoder and decoder. The
encoder takes in an image as input and extracts as many Fig 2:- System Architecture
features from the position of the target. It then consolidates
all features from n images given into the GAN network as The images are first preprocessed. They are aligned
f(x). The decoder of the GAN generator network takes this and the appropriate resolution is set to train them to the
feature set and restores the image of the target with machine. Images are then vectorised. Their attributes that
different poses denoted by different and unique pose-codes. are distinguishable are stored as features into a feature set.
This image result is then fed to the discriminator and the The features are then given as input to the training GAN for
value is generated if the target is a real person. Also, the further training procedure. The images that this machine
value is generated stating the identity of the person on the works with would be the normal and practical resolution
training database. The generator and the discriminator were that would be clear enough to distinguish features and
first introduced as multiple neural networks but over time, identify human faces.
it was found that the number of weights introduced in every
layer of the multiple neural networks were infeasible with
the increase of the number of pixels as the resolution if the
training dataset images increased.

IJISRT20JUN1116 www.ijisrt.com 1527

Volume 5, Issue 6, June – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
ACKNOWLEDGEMENT

We would like to express our immense gratitude and

profound thanks to all those who helped us to make this
project a great success. We express our gratitude to the
almighty God for all the blessings endowed on us. We
express my thanks to our Executive Director REV.FR.
GEORGE PAREMAN, Director DR.ELIZABETH ELIAS,
Principal DR. NIXON KURUVILA for providing us with
such a great opportunity. We are thankful for the help and
appreciation we received from the head of the department
DR. RAJESWARI M, project coordinators MS. DEEPA
DEVASSY, MS. ANLY ANTONY and MR. WILLSON
Fig 3:- Use Case Diagram JOSEPH We would also extend my deep sense of gratitude
to our project guide MS. ANUSREE K for her guidance
The advantages are: and advice. We would like to express our gratitude towards
 Easy recognition my parents for their timely cooperation and encouragement.
 Accommodates ambiguous input faces Every project is successful due to the effort of many
 Efficient with increased use people. Our thanks and appreciation go to all our peers who
 Fast and accurate results have given us their valuable advice and support and pushed
 Scope for broader applications us into successfully completing this project.

IV. RESULTS REFERENCES

Generative Adversarial Networks are used to identify [1]. D. J. Robertson, R. S. Kramer, and A. M. Burton,
targets or criminals even with slightly different facial ``Fraudulent ID using face morphs: Experiments on
features, including hair colour, eye colour, facial hair human and automatic recognition,'' PLoS ONE, vol.
structure, etc. Thus, even when the target can disguise 12, no. 3, 2017, Art. no. e0173319.
themselves, they are identified using the adversarial [2]. J. Zhao, L. Xiong, P. K. Jayashree, J. Li, F. Zhao,
technique of machine learning. Generators can generate Z. Wang, P. S. Pranata, P. S. Shen, S. Yan, and J.
more realistically fake images with slightly tuned features Feng. Dual-agent gans for photorealistic and Identity
which the discriminator will extract features from and can preserving profile face synthesis. In NIPS, 2017.
further generate the corresponding identity code of the [3]. Karras, Tero, et al. "Progressive growing of GANS
target from the database. Thus face recognition is carried for improved quality, Stability, and variation," arXiv
out successfully. preprint arXiv:1710.10196, 2017.
[4]. T. Salimans, I. Goodfellow, W. Zaremba, V.
V. CONCLUSION Cheung, A. Radford, and X.Chen, “Improved
Techniques for Training GANs,” in Advances In
A system that can recover undisguised faces could be Neural Information Processing Systems, 2016, pp.
helpful for criminal investigation. In particular, witnesses 2234–2242.
should be able to make use of these processed images to [5]. X. Yu and F. Porikli, “Ultra-resolving face images
identify the criminals among a series of ID photos, which by discriminative generative networks,” in Proc.
typically include no disguise or in person among a number European Conf. Comput. Vis. Springer, 2016, pp.
of held suspects. People utilizing online dating apps could 318–333.
also utilize the system to reverse the real person behind the [6]. C. Peng, X. Gao, N. Wang, and J. Li, ``Face
facial disguise, a feature that many find useful. This project recognition from multiple stylistic sketches:
is built on current work related to GAN-based style Scenarios, datasets, and evaluation,'' Pattern
transform methods that are commonly employed for Recognit., vol. 84, pp. 262272, Dec. 2018.
applying facial disguise. There have been significant [7]. M. S. Sarfraz and R. Stiefelhagen, “Deep perceptual
improvements in machine capability to identify and verify mapping for cross modal face recognition,”
human faces over the past few years. Device makers are International Journal of Computer Vision, vol. 122,
already taking advantage of such development by no. 3, pp. 426–438, May 2017.
equipping their latest phones and tablets with AI co- [8]. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
processor and powerful image processing algorithms. D. Warde-Farley, S. Ozair, A. Courville, and Y.
However, the recent trend has mostly focused on making Bengio, “Generative adversarial nets,” in Advances in
facial identification and verification invariant to facial neural information processing systems, 2014, pp.
features. These works certainly help machines recognize 2672– 26.
human faces. However, most humans are interested in
seeing people in the natural state without any facial
disguise.

IJISRT20JUN1116 www.ijisrt.com 1528