An adaptive algorithm based on principal component analysis- deep learning for anomalous events detection

Abstract

One of the most often used applications of human activity detection is anomaly detection, which is covered in this paper. Providing security for a person is a key issue in every community nowadays because of the constantly expanding activities that pose danger, from planned violence to harm caused by an accident. Existing classical closed-circuit television considered is insufficient since it needs a person to stay awake and constantly monitor the cameras, which is expensive. In addition, a person's attention decreases after a certain time. For these reasons, the development of an automated security system that can identify suspicious activities in real-time and quickly aid victims is required. Because identifying activity must be with high accuracy, and in the shortest possible time. We adopt an adaptive algorithm based on the combination of machine learning (ML), principal component analysis (PCA) and deep learning (DL). The UCF-crime dataset was used for the experimentation in this work. Where the area under the curve (AUC) with the proposed approach was equal to 94.21% while the detection accuracy was equal to 88.46% on the test set database. The suggested system has demonstrated its robustness and accomplishment of the best accuracy when compared with earlier designed systems.

Full text

Indonesian Journal of Electrical Engineering and Computer Science
Vol. 29, No. 1, January 2023, pp. 421~430
ISSN: 2502-4752, DOI: 10.11591/ijeecs.v29.i1.pp421-430  421
Journal homepage: http://ijeecs.iaescore.com
An adaptive algorithm based on principal component analysis-
deep learning for anomalous events detection
Zainab K. Abbas, Ayad A. Al-Ani
Department of Information and Communication Engineering, College of Information Engineering, Al-Nahrain University,
Baghdad, Iraq
Article Info
ABSTRACT
Article history:
Received Aug 13, 2022
Revised Sep 9, 2022
Accepted Sep 19, 2022
One of the most often used applications of human activity detection is
anomaly detection, which is covered in this paper. Providing security for a
person is a key issue in every community nowadays because of the constantly
expanding activities that pose danger, from planned violence to harm caused
by an accident. Existing classical closed-circuit television considered is
insufficient since it needs a person to stay awake and constantly monitor the
cameras, which is expensive. In addition, a person's attention decreases after
a certain time. For these reasons, the development of an automated security
system that can identify suspicious activities in real-time and quickly aid
victims is required. Because identifying activity must be with high accuracy,
and in the shortest possible time. We adopt an adaptive algorithm based on
the combination of machine learning (ML), principal component analysis
(PCA) and deep learning (DL). The UCF-crime dataset was used for the
experimentation in this work. Where the area under the curve (AUC) with the
proposed approach was equal to 94.21% while the detection accuracy was
equal to 88.46% on the test set database. The suggested system has
demonstrated its robustness and accomplishment of the best accuracy when
compared with earlier designed systems.
Keywords:
Anomaly detection
Bidirectional long short term
memory
Deep learning
Machine learning
Principal component analysis
Resnet50
This is an open access article under the CC BY-SA license.
Corresponding Author:
Zainab K. Abbas
Department of Information and Communication Engineering, College of Information Engineering
Al-Nahrain University, Baghdad, Iraq
Email: zainabkudair@gmail.com
1. INTRODUCTION
Video surveillance systems (VSS) are frequently utilized to increase public safety in various settings,
including malls, hospitals, banks, markets, intelligent cities, educational institutions, and roadways [1]. The
main goal of security applications is typically the accuracy and speed of video anomaly identification [2]. For
the sake of public safety, numerous surveillance cameras have recently been put in numerous areas throughout
the world [3]. These cameras continuously generate massive volumes of video data [3]. Therefore, real-time
video analysis and finding abnormal cases require a lot of human resources. In addition, it is subject to fault
due to human attention loss over time [1]. Because human monitoring is ineffective, automatic anomaly
detection solutions based on artificial intelligence (AI) algorithms become necessary in surveillance systems [4].
Various techniques in the literature identify anomaly actions as "the occurrence of variance in regular
patterns" [3]. Traffic security, automated intelligent visual monitoring, and crime prevention are some
applications for abnormal event detection in surveillance videos [5]. Due to the lack of actual anomalous
instances, video anomalous detection was formerly assumed to be a one-class classification task [6]-[8] that is
to mean the classifier model is trained on normal movies, and a video is classified as anomalous when irregular

 ISSN: 2502-4752
Indonesian J Elec Eng & Comp Sci, Vol. 29, No. 1, January 2023: 421-430
422
patterns are seen in the testing [5]. As a result, various normal behaviors may swerve from normal events in
the training set, resulting in false alarms [1], [9].
Different studies were performed in this field, and all these strategies were adopted for a specific
situation. A study performed by Waqas et al. [10] provided a framework that can recognize abnormal attitudes
and tell the user of the type of behavior. Another study suggests adaptively compressing each video before to
being transmitted to the event detection system by Shreyas et al. [11]. Anala et al. [12], Hao et al. [13] and
Dubey et al. [14] regarded the detect anomalous behavior as a regression problem. Another study offered a
lightweight convolutional neural network (CNN) by Ullah et al. [5]. Ullah et al. [3] proposed an intelligent
anomaly detection system that depends on combining the ResNet50 with multilayer bidirectional long short
term memory (BiLSTM). Zaheer et al. [15] presented a weakly supervised anomaly detection method that
trains the model using video-level labels. Another method for handling anomaly detection and classification
using a weakly supervised learning model was provided by Majhi et al. [16]. Employing multi-detail ideas in
both the temporal and spatial dimensions as input, a dual branch network has been developed by Wu et al. [17].
To detect video anomalies, Cao et al. [18] suggested taking into consideration the spatial-temporal relationships
between video parts. Abbas and Al-Ani [2] suggest compressing each video using high-efficiency video coding
(H265) before feeding the video into the anomaly detection systems. An algorithm for reducing the size of the
extracted features has been suggested by Abbas and Al-Ani before anomaly identification [19].
Even though the word "anomaly" is used frequently in literature, there isn't a standard definition of it
yet [9]. The most of existing technologies have a high rate of false alarms. Additionally, the effectiveness of
these approaches is reduced when used in real-world situations, even though they work effectively on
basic databases.
To overcome these difficulties, we suggested reducing the dimensionality of the data features
extracted using pre-trained convolution neural networks (Resnet50), by using principal component analysis
(PCA) [20] to improve the performance of the model and reduce the model complexity. After that, we feed the
features to our classifier model, which is a BiLSTM. We employ a weakly supervised method based on spatio-
temporal features and BiLSTM to train our classifier model. When the context of the input is needed, BiLSTMs
have proven to be highly helpful. While in a unidirectional LSTM, information moves from backward to
forward, the BiLSTM uses two hidden states to flow information not only backward to forward, but also
forward to backward. As a result, BiLSTMs are better able to perceive the context [21].
2. METHODS
The results of a study that compared the various methods used for anomaly detection showed that
deep learning (DL) has outperformed other methods in this field [1]. The recommended approach used in this
work is split into three stages:
− Feature extraction using Resnet50.
− Dimensionality reduction using PCA.
− Anomaly events detection using BiLSTM.
This work focuses on the assessment of the mixture of the machine learning (ML) and DL algorithms
for anomaly event detection purposes in videos for the first time on the UCF-Crime database. Figure 1
illustrates the suggested framework.
Figure 1. The suggested framework

Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 
An adaptive algorithm based on principal component analysis-deep learning for … (Zainab K. Abbas)
423
2.1. Input database UCF-crime
The UCF-Crime database is the database that was used in this work [22], [23]. This comprises 13 different
categories of anomalies, such as explosions, fights, abuse, and accidents, in addition to normal events. The collection
contains 1900 surveillance videos, with almost equal numbers of normal and abnormal videos. The training set
included 810 anomalous and 800 normal videos, while the testing set included 140 anomalous and 150 normal videos
[10], [24]. This set includes almost 129 hours of videos with a resolution of 320x240 and 13 million frames, with
different videos length [10], [22], [24]. We selected this database because it includes a variety of anomalous event
categories since the irregularities in it have a big influence on public safety. However, there are two issues with this
database. The first is that this dataset's anomalous class has high inter-class variations. The second, video-level labels,
which mean we only know that each movie has an abnormality but not which exact part is abnormal. This maybe
leads to overfitting [2], [22]. In this work, we picked the video with a length less than or equal to 2 min, depending
on this condition, we had 1324 videos in total divided as follows: 1116 videos for the training purpose (90% for
training and 10% for validation), and the remainder, which was 208, was used for the testing purpose.
2.2. Machine learning (ML) and deep learning (DL)
It is possible to successfully handle unstructured data with DL which is considered a subset of ML.
DL approaches exceed current ML approaches. It makes it possible for computational models to gradually
learn and comprehend the features of the information at various levels. DL became more and more prevalent
as data availability expanded and powerful computers. The DL method transforms the input into levels, each
of which can extract features and send them to the next layer. Initial layers gather basic data, which is then
integrated with later layers to provide a complete description. DL can be accomplished using a variety of
designs, including CNN, pre-trained networks, recurrent neural networks (RNN), and others. The efficiency of
DL classifiers greatly increases as the number of data increases when compared to traditional learning methods.
DL algorithms perform better as the amount of training data increases, but traditional machine learning
algorithms' performance stabilizes after a certain amount of training data. Although deep structures require
more time to train, they perform better than straightforward artificial neural networks (ANNs). However,
strategies like transfer learning and GPU computing can shorten the training period [1], [2], [25].
One of the neural network kinds is CNN which contain convolutional layers. Despite the processing
of the spatial data using CNN being good, for handling the sequential data RNN is better. This is due to RNN
utilizing state variables to save the past data and use them beside the present inputs to determine the present
outputs [2], [26]. Most of the time, CNN is used in image processing. Various items in the image are given
biases and weights, and this separates them. CNN needs lower preparation than other classification techniques.
Where it uses the appropriate filters to extract both temporal and spatial links in an image. Some CNN kinds
are ZFNet, VGGNet, ResNet, GoogleNet, AlexNet and LeNet [2], [25].
RNNs, on the other hand, use prior outputs as inputs to determine the condition of the current situation.
Knowledge can be remembered by RNN's hidden layers. To modify the hidden state, the output generated in
the previous state was applied. Due to their ability to remember previous inputs, RNNs can be utilized to
estimate time series. An example of an RNN is the LSTM [2], [25].
2.2.1. Features extraction
The 50-layer deep convolutional neural network ResNet50, which has 23.5 million learnable
parameters, was utilized in this study to extract features [3]. This network generates 1000 characteristics for
each frame [2]. ResNet50's input is 224 by 224 pixels in size. For this reason, in this research, the longest edges
of a movie were cropped and resized to match the input dimensions using a center crop.
2.2.2. Dimensionality reduction - principal component analysis (PCA)
This technique can be found in the literature with different names like the singular value
decomposition (SVD), hotelling transform, the Karhunen Loeve transform (KLT), and the empirical orthogonal
function (EOF) method [27]. It is an old, straightforward statistical approach that changes data from a higher
dimension to a lower dimension in the hopes of gaining a better understanding of the data. For data
compression, dimensionality reduction, and data visualization, PCA is utilized. In the case of dimensionality
reduction, it greatly simplifies the problem by drastically lowering the number of features. If the original dataset
has x features, the new dataset will have y features, where y is less than x. Some information will be lost due
to the new dimension being smaller than the original one. PCA can aid in the discovery of factors buried deep
within the data. The principal components are uncorrelated, linear combinations of the original data. 90% of
the original signal could be accounted for by two or three of the main components. In machine learning, PCA
is an unsupervised learning approach for dimensionality reduction. It is a statistical technique that converts a
set of linearly uncorrelated features from correlated feature observations via orthogonal transformation. The
principal components (PC) indicate these newly modified features [20], [27], [28]. Figure 2 shows the steps of
PCA applied in this work [27].

 ISSN: 2502-4752
Indonesian J Elec Eng & Comp Sci, Vol. 29, No. 1, January 2023: 421-430
424
Figure 2. The PCA steps
2.2.3. Classifier
Unlike previous studies that fed the classifier model with features extracted from the features
extraction stage, in this work, we fed the classifier model with features got from the dimensional reduction
stage. The BiLSTM consider suitable for long sequential data such as videos [2]. So, in this research, we used
the BiLSTM as a classifier model for anomaly detection.
3. RESULTS AND DISCUSSION
The MATLAB software environment (version 2021a) was utilized to implement the computer codes
in this research. The work was done on the computer with the following specification: Windows 10, Intel Core
i7 processor, 1 TB SSD hard drive, 64-bit operating system, 16 GB RAM, and NVIDIA GeForce MX450
graphics processing unit. The results of each stage were as follows:
3.1. Input dataset UCF-crime
The UCF-Crime database was used for this study's experiments. The 13 anomaly classes in addition
to the normal class were used. Since the anomaly identification in this work is done at the video level, the
duration of the video had no bearing on the functioning of anomaly detection, so we chose videos that were no
longer than or equal to two minutes, depending on this condition, we had 1324 videos in total divided as
follows: 1116 videos for the training purpose (90% for training and 10% for validation), and the remainder,
which was 208, was used for the testing purpose.
3.2. Machine learning (ML) and deep learning (DL)
3.2.1. Features extraction
In this study, feature extraction was accomplished using a pre-trained model called ResNet50. Instead
of the SoftMax layer, the fc1000 layer is used to extract the features. This means each frame in the video after
this step will be represented by 1000 features, i.e. for a video with (x) number of frames the number of features

Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 
An adaptive algorithm based on principal component analysis-deep learning for … (Zainab K. Abbas)
425
will be (1000 * x) and this is considered a huge data, for this reason, in this paper, we suggest for the first time
feed the extracted features to PCA model before classifying them, to reduce the training time in the same time
increase the accuracy of the classification.
3.2.2. Dimensionality reduction - principal component analysis (PCA)
In this stage, PC is calculated for the extracted features where PC refers to these newly modified features
which will be fed the classifier model instead of the extracted features from ResNet50. In this work, the PC for
different variance values has been calculated for each video in the database. Figures 3 and 4 show the PC numbers
for all videos at different variance values for train and test data, respectively, for the UCF-Crime dataset. Where
the x-axis represents the video number, the y-axis represents the PC number of the video.
Figure 3. The video’s PC number for train datasets at different variance values
Figure 4. The video’s PC number for test datasets at different variance values

 ISSN: 2502-4752
Indonesian J Elec Eng & Comp Sci, Vol. 29, No. 1, January 2023: 421-430
426
3.2.3. Classifier
Application of decision classifiers becomes necessary after the features are complete. In this work,
the BiLSTM classification model has been used. The parameters of the classifier model are shown in Table 1
for different variance values, Adam optimizer was utilized for all cases. The value of L2Regularization and
dropout was selected to lessen the model's overfitting. The model parameters were selected by trial and error.
Table 1. The parameter values of the classifier model at different variance values
Classifier
Parameters
Variance value
99%
95%
90%
85%
80%
75%
70%
65%
60%
55%
Minimum Batch Size
32
32
32
32
32
32
32
32
32
32
Hidden layer nodes No.
112
112
112
112
112
112
112
112
112
112
Dropout
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
Initial Learning Rate
1e-5
1e-5
1e-5
1e-5
1e-5
1e-5
1e-5
1e-5
1e-5
1e-5
Maximum epochs
170
170
190
170
170
170
170
170
170
170
L2Regularization
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
Following the previous research, the area under curve (AUC) and the receiver operating characteristics
(ROC) were used as performance measures to assess the performance of the suggested work. Moreover,
we determine the detection accuracy of our classifier model. The classifier accuracy was determined using
the (1) [29].
   
 x 100% (1)
The experimental results show the efficiency of our proposed framework, as it detects anomalous
events with greater precision than existing methods. Our classifier model's ROC curve is shown on the left side
in Figures 5-14, while the confusion matrix is shown on the right, for different variance values. It is clear that
the highest value for AUC was recorded at a variance value equal to 90%, while the highest value for detection
accuracy was recorded at a variance value equal to 95% where the true positive (TP) which means an anomaly
classified as anomaly was equal to 84, false negative (FN) which means an anomaly classified as normal was
equal to 9, true negative (TN) which means normal classified as normal was equal to 100, and false positive
(FP) which means normal classified as an anomaly was equal to 15.
AUC = 92.30%
Detection accuracy = 84.13%
Figure 5. The classifier model ROC curve and the confusion matrix for variance = 99%
AUC = 93.60%
Detection accuracy = 88.46%
Figure 6. The classifier model ROC curve and the confusion matrix for variance = 95%

Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 
An adaptive algorithm based on principal component analysis-deep learning for … (Zainab K. Abbas)
427
AUC = 94.21%
Detection accuracy = 86.54%
Figure 7. The classifier model ROC curve and the confusion matrix for variance = 90%
AUC = 94.11%
Detection accuracy = 87.50%
Figure 8. The classifier model ROC curve and the confusion matrix for variance = 85%
AUC = 93.61%
Detection accuracy = 86.54%
Figure 9. The classifier model ROC curve and the confusion matrix for variance = 80%
AUC = 93.28%
Detection accuracy = 87.50%
Figure 10. The classifier model ROC curve and the confusion matrix for variance = 75%

 ISSN: 2502-4752
Indonesian J Elec Eng & Comp Sci, Vol. 29, No. 1, January 2023: 421-430
428
AUC = 92.88%
Detection accuracy = 86.54%
Figure 11. The classifier model ROC curve and the confusion matrix for variance = 70%
AUC = 92.13%
Detection accuracy = 85.58%
Figure 12. The classifier model ROC curve and the confusion matrix for variance = 65%
AUC = 93.25%
Detection accuracy = 86.54%
Figure 13. The classifier model ROC curve and the confusion matrix for variance = 60%
AUC = 93.44%
Detection accuracy = 86.06%
Figure 14. The classifier model ROC curve and the confusion matrix for variance = 55%

Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 
An adaptive algorithm based on principal component analysis-deep learning for … (Zainab K. Abbas)
429
After explaining the suggested system details, it is vital to compare the suggested system with the
previous studies. The AUC scores have been compared with the previous research works in Table 2, and it is
clear that the proposed system achieved the highest AUC of 94.21%.
Table 2. AUC score comparison between the proposed work and the previous works
Method
AUC %
Waqas et al. [10]
75.41
Anala et al. [12]
85
Shreyas et al. [11]
79.8
Hao et al. [13]
81.22
Dubey et al. [14]
81.91
Ullah et al. [5]
78.43
Ullah et al. [3]
85.53
Zaheer et al. [15]
78.27
Majhi et al. [16]
82.12
Wu et al.[17]
87.65
Cao et al. [18]
83.14
Abbas and Al-Ani [2]
90.16
Abbas and Al-Ani [19]
93.61
Our Adaptive Algorithm
94.21
6. CONCLUSION
The proposed anomalous event detection system based on a combination between ML and DL is
designed and tested, in which the feature vector for each video was extracted using pre-trained Resnet50. After
that, the feature reduction algorithm PCA was used to remove particular features, which was used for the first
time in such work. Finally, the new feature vectors were fed into BiLSTM for abnormal and normal class
detection. In comparison with previous works on anomalous detection approaches, the suggested system has
been shown to have superior accuracy. According to the experimental results, the AUC value for the UCF-
Crime database increased by up to 94.21%. We also measured the classifier's detection accuracy, which came
out to be 88.46%. And this demonstrates the effectiveness of our suggestion to increase the accuracy of
anomaly event detection, with the least amount of both negative and positive false alarms. Future work will
focus on investigating different feature extraction models, feature selection techniques, and dimensionality
reduction techniques to combine them with our proposed system to improve the accuracy indicator.
REFERENCES
[1] Z. K. Abbas and A. A. Al-Ani, “A comprehensive review for video anomaly detection on videos,” Proceedings of the 2nd 2022
International Conference on Computer Science and Software Engineering, CSASE. 2022, pp. 30–35, doi:
10.1109/CSASE51777.2022.9759598.
[2] Z. K. Abbas and A. A. Al-ani, “Anomaly detection in surveillance videos based on H265 and deep learning,” Int. J. Adv. Technol.
Eng. Explor., vol. 9, no. 92, pp. 910–922, 2022, doi: 10.19101/IJATEE.2021.875907.
[3] W. Ullah, A. Ullah, I. U. Haq, K. Muhammad, M. Sajjad, and S. W. Baik, “CNN features with bi-directional LSTM for real-time
anomaly detection in surveillance networks,” Multimed. Tools Appl., vol. 80, no. 11, pp. 16979–16995, 2021, doi: 10.1007/s11042-
020-09406-3.
[4] H. İ. Öztürk and A. B. Can, “ADNet: temporal anomaly detection in surveillance videos,” Lect. Notes Comput. Sci. (including
Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 12664 LNCS, pp. 88–101, 2021, doi: 10.1007/978-3-030-68799-17.
[5] W. Ullah, A. Ullah, T. Hussain, Z. A. Khan, and S. W. Baik, “An efficient anomaly recognition framework using an attention
residual lstm in surveillance videos,” Sensors, vol. 21, no. 8, 2021, doi: 10.3390/s21082811.
[6] R. Morais, V. Le, T. Tran, B. Saha, M. Mansour, and S. Venkatesh, “Learning regularity in skeleton trajectories for anomaly detection
in videos,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 2019, pp. 11988–11996, doi: 10.1109/CVPR.2019.01227.
[7] Y. Fan, G. Wen, D. Li, S. Qiu, M. D. Levine, and F. Xiao, “Video anomaly detection and localization via Gaussian mixture fully
convolutional variational autoencoder,” Comput. Vis. Image Underst., vol. 195, 2020, doi: 10.1016/j.cviu.2020.102920.
[8] F. P. dos Santos, L. S. F. Ribeiro, and M. A. Ponti, “Generalization of feature embeddings transferred from different video anomaly
detection domains,” J. Vis. Commun. Image Represent., vol. 60, pp. 407–416, 2019, doi: 10.1016/j.jvcir.2019.02.035.
[9] J. Ren, F. Xia, Y. Liu, and I. Lee, “Deep video anomaly detection : Opportunities and challenges,” In 2021 International Conference
on Data Mining Workshops (ICDMW), pp. 959-966, IEEE, 2021, doi: 10.1109/ICDMW53433.2021.00125.
[10] S. Waqas, C. Chen, and S. Mubarak, “Real-world anomaly detection in surveillance videos,” Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 6479-6488, 2019, doi: 10.1109/HiPCW.2019.00031.
[11] D. G. Shreyas, S. Raksha, and B. G. Prasad, “Implementation of an anomalous human activity recognition system,” SN Comput.
Sci., vol. 1, no. 3, 2020, doi: 10.1007/s42979-020-00169-0.
[12] M. R. Anala, M. Makker, and A. Ashok, “Anomaly detection in surveillance videos,” Proc. - 26th IEEE Int. Conf. High Perform.
Comput. Work. HiPCW, 2019, pp. 93–98, doi: 10.1109/HiPCW.2019.00031.
[13] W. Hao et al., “Anomaly event detection in security surveillance using two-stream based model,” Hindawi Secur. Commun.
Networks, vol. 2020, 2020, doi: 10.1155/2020/8876056.

 ISSN: 2502-4752
Indonesian J Elec Eng & Comp Sci, Vol. 29, No. 1, January 2023: 421-430
430
[14] S. Dubey, A. Boragule, J. Gwak, and M. Jeon, “Anomalous event recognition in videos based on joint learning of motion and
appearance with multiple ranking measures,” Appl. Sci., vol. 11, no. 3, pp. 1–21, 2021, doi: 10.3390/app11031344.
[15] M. Z. Zaheer, J. Lee, M. Astrid, A. Mahmood, and S.-I. Lee, “Cleaning label noise with clusters for minimally supervised anomaly
detection,” arXiv e-prints, pp. 3–6, 2021, [Online]. Available: http://arxiv.org/abs/2104.14770.
[16] S. Majhi, S. Das, F. Bremond, R. Dash, and P. K. Sa, “Weakly-supervised joint anomaly detection and classification,” 16th IEEE
International Conference on Automatic Face and Gesture Recognition (FG 2021), pp. 1-7. IEEE, 2021 2021, doi:
10.1109/FG52635.2021.9667006.
[17] J. Wu et al., “Weakly-supervised spatio-temporal anomaly detection in surveillance video,” arXiv preprint arXiv:2108.03825,
pp. 1172–1178, 2021, doi: 10.24963/ijcai.2021/162.
[18] C. Cao, X. Zhang, S. Zhang, P. Wang, and Y. Zhang, “Adaptive graph convolutional networks for weakly supervised anomaly
detection in videos,” arXiv:2202.06503, 2022.
[19] Z. K. Abbas and A. A. Al-Ani, “Detection of anomalous events based on deep learning-BiLSTM,” Inpress.
[20] H. Alquran et al., “Cervical cancer classification using combined machine learning and deep learning approach,” Comput. Mater.
Contin., vol. 72, no. 3, pp. 5117–5134, 2022, doi: 10.32604/cmc.2022.025692.
[21] A. A. Sharfuddin, M. N. Tihami, and M. S. Islam, “A deep recurrent neural network with bilstm model for sentiment classification,”
2018 Int. Conf. Bangla Speech Lang. Process. ICBSLP 2018, no. September, 2018, doi: 10.1109/ICBSLP.2018.8554396.
[22] R. Maqsood, U. I. Bajwa, G. Saleem, R. H. Raza, and M. W. Anwar, “Anomaly recognition from surveillance videos using 3D
convolution neural network,” Multimed. Tools Appl., vol. 80, no. 12, pp. 18693–18716, 2021, doi: 10.1007/s11042-021-10570-3.
[23] “UCF-Crime-database,” [Online]. Available: https://www.dropbox.com/sh/75v5ehq4cdg5g5g/AABvnJSwZI7zXb8_myBA0CLHa?dl=0.
[24] M. S. P. Kumari and A. K. Bedi, “Multimedia datasets for anomaly detection: A review.” arxiv:2112.05410, 2022.
[25] A. Mathew, P. Amudha, and S. Sivakumari, “Deep learning techniques: an overview,” Adv. Intell. Syst. Comput., vol. 1141,
no. August 2020, pp. 599–608, 2021, doi: 10.1007/978-981-15-3383-9_54.
[26] J. M. Czum, “Dive into deep learning,” J. Am. Coll. Radiol., vol. 17, no. 5, pp. 637–638, 2020, doi: 10.1016/j.jacr.2020.02.005.
[27] A. A. Al-Ani, L. A. Al-Ani, and Y. A. Jasim, “Iris recognition using independent and principal component analysis,” Dissertation,
2018.
[28] I. T. Jollife and J. Cadima, “Principal component analysis: A review and recent developments,” Philos. Trans. R. Soc. A Math. Phys.
Eng. Sci., vol. 374, no. 2065, 2016, doi: 10.1098/rsta.2015.0202.
[29] S. H. Abd, I. A. Hashim, and A. S. A. Jalal, “Automatic deception detection system based on hybrid feature extraction techniques,”
Indones. J. Electr. Eng. Comput. Sci., vol. 26, no. 1, pp. 381–393, 2022, doi: 10.11591/ijeecs.v26.i1.pp381-393.
BIOGRAPHIES OF AUTHORS
Zainab K. Abbas Born in Baghdad/Iraq, in 1993. She gained B.Sc. and M.Sc.,
from Dept. of Control and system engineering/University of technology, in 2015 and 2018,
respectively. She is currently a Ph.D. student in the Department of Information and
Communication Engineering, College of Information Engineering, Al-Nahrain University,
Baghdad, Iraq. Her works studies are focused on Digital processing with artificial
intelligence. She can be contacted at email: zainabkudair@gmail.com.
Ayad A. Al-Ani he is born in Baghdad/Iraq, in 1961. He gained the B.Sc., M.Sc.,
and Ph.D., from Dept. of Physics/College of Science/Baghdad University, in 1983, 1990,
and 1995, respectively. Previously, he was University. Vice President for Administrative
affairs/Al-Nahrain Univ., Deputy Dean for Post Graduate Studies and Scientific
Research/Baghdad Univ, and Head of Space and Astronomy Dept./College of
Science/Baghdad University. Since 2006, he is a Professor of Digital Image Processing in
Dept. of Information and Communication Engineering/College of Information
Engineering/Al-Nahrain University. He published 74 papers and 4 books. He can be
contacted at email: ayad.a@nahrainuniv.edu.iq.