Page 10 - IJEEE-2022-Vol18-ISSUE-1
P. 10
6| Kadhim & Al-Darraji
D. Classification 2) Prediction Model for Face Images
This paper uses the CNN network to deal with the Recognizing and confirming individuals from an
original images and to distorted images of the dataset. It image of their face is a computer vision job known as face
enhances the results with high accuracy. This network helps recognition. Although various open-source implementations
to reduce or remove perturbation from the original images. and pre-trained models for Google's net facial recognition
However, building a CNN structure consists of ten layers, system about the face, face image predictions are being used
made up of five “convolutional layers” and five “pooling in increasing numbers by facial analysis apps, due to the fact
layers” with “Fully-Connected (FC) layers”. After that, they that the technology has a wide range of applications.
extract all the features and use the Softmax classifier for However, although the existing models are still lacking in
recognition. Then, all the original and distorted images are accuracy, they are hindered by the vast variety of face
trained with several methods of tests on the neural network. images that exist (such as differences in lighting, poses, and
Thus, Figure 10 shows the structure of CNN and Figure 11 angles). It is necessary to follow such a process in order to
shows the general structure of CNN respectively. The use these models in real-world situations.
classification process is composed of two steps, which are
the training and prediction model for face images. For the accurate prediction of a collection of face
images, an improved deep learning structure based on the
Fig. 10: Structure of the proposed convolutional neural combination of attention and residual convolutional
network. networks was presented.
Using multitasking learning, the accuracy of face
prediction may be enhanced by adding predicted faces to the
feature embedding of the face classifier, which can then be
used to further train the model. When our proposed model
was trained, an image of a well-known individual and a
frequently used dataset were used, and the results were very
remarkable. Observing our trained model's attention maps, it
can be seen that it has learned to be aware of different facial
areas over time. Image prediction is accomplished via the
usage of the CNN. The methods used for image processing
are shown in the following Figure 12.
Fig. 11: The general structure of CNN of the original and Fig. 12: The image prediction process.
distorting of the faces recognition system.
V. EXPERIMENTAL AND RESULTS
1) Training
The images are chosen from the LFW database and Experiments were performed to illustrate the suggested
method's efficacy of recognizing faces in spite of different
then the number of classes is determined, which is about 10 attack generations on the CNN model.
classes for males and females in different positions. The total
number of original images is about 1,083 images. Then these A. Setup
images are divided randomly and manually into 70% (720
images) for training and 30% (363 images) for testing. Databases: LFW (Labeled Faces Wild) [19] is used
for the tests. Face images are in the LFW database for the
After that, all database images are distorted with three purpose of studying the issue of unlimited face recognition.
kinds of attacks; FGSM, Deep Fool, and PGD. Finally, the To run the tests, the LFW database will be used. However,
CNN model will be trained using original and distorted the method's performance is assessed using the LFW
images set to get high accuracy of recognition which is database. The database has more than 13,000 images of
strong to face any problems on the FR system. faces. Each portrait is labeled with the subject's name. The
images are available in two sizes (250 by 250) pixels with
variations in emotion, posture, time, and gender. However,
not all classes were used because not all classes have a large
number of images, some of them with only one or two
images, which is not sufficient to recognize faces. Therefore,
10 classes were selected, in which each person has at least 50
