|
|
Direction-of-arrival estimation for conventional
co-prime arrays using probabilistic Bayesian neural
networks
Wael Elshennawy, Orange Business Services, Egypt
Abstract:
The paper investigates the direction-of-arrival (DOA) estimation of narrow
band signals with conventional co-prime arrays by using efficient
probabilistic Bayesian neural networks (PBNN). A super resolution DOA
estimation method based on Bayesian neural networks and a spatially
overcomplete array output formulation overcomes the pre-assumption
dependencies of the model-driven DOA estimation methods. The proposed
DOA estimation method utilizes a PBNN model to capture both
data and model uncertainty. The developed PBNN model is trained to
do the mapping from the pseudo-spectrum to the super resolution spectrum.
This learning-based method enhances the generalization of untrained
scenarios, and it provides robustness to non-ideal conditions, e.g.,
small angle separation, data scarcity, and imperfect arrays, etc. Simulation
results demonstrate the root mean square error (RMSE) and loss
curves of the PBNN model in comparison with deterministic model and
spatial-smoothing MUSIC (SS-MUSIC) method. The proposed Bayesian
estimator improves the DOA estimation performance for the case of low
signal-to-noise ratio (SNR) or with a limited number of model trainable
variables or spatially adjacent signals.
Full Paper (in PDF)
|
Hybrid Deep Learning for Assembly Action Recognition in Smart Manufacturing
Abdul Matin, University of Technology Sydney, Australia
Md Rafiqul Islam, University of Technology Sydney, Australia
Yeqian Zhu, University of Technology Sydney, Australia
Xianzhi Wang, University of Technology Sydney, Australia
Huan Huo, University of Technology Sydney, Australia
Guandong Xu, University of Technology Sydney, Australia
Abstract:
Deep learning algorithms have become essential in assembly action recognition
(AAR) for driving advancements in intelligent manufacturing.
While numerous sensor systems and algorithms are developing, their
real-world applicability and robustness within the manufacturing sector
need validation. Artificial intelligence (AI) applications in manufacturing
have gained significant attraction in both academic and industrial
circles. One key aspect of future smart manufacturing is identifying the
actions of manufacturing workers, particularly monitoring repetitive assembly
tasks, to guide them and improve efficiency. This recognition
facilitates real-time efficiency measurement and evaluation of workers
while providing augmented reality instructions to enhance their performance
on the job. This paper introduces a hybrid deep-learning approach
combining 3D CNN and ConvLSTM2D models to monitor assembly
tasks to recognize human actions within the manufacturing context.
The model’s performance is evaluated through simulations conducted
on the HA4M dataset, comprising diverse multimodal data-capturing
actions executed by various individuals constructing an epicyclic gear
train (EGT). The proposed hybrid model demonstrated superior performance
on the HA4M dataset relative to baselines.
Full Paper (in PDF)
|
Automated Fracture Detection from Pelvic X-ray:
The Impact of Appropriate Labeling on the
Performance of Deep Convolutional Neural Network
Rashedur Rahman, University of Hyogo, Japan
Naomi Yagi, University of Hyogo, Japan
Keigo Hayashi, Hyogo Prefectural Harima-Himeji General Medical Center, Japan
Akihiro Maruo, Hyogo Prefectural Harima-Himeji General Medical Center, Japan
Hirotsugu Muratsu, Hyogo Prefectural Harima-Himeji General Medical Center, Japan
Sayoji Kobashi, University of Hyogo, Japan
Abstract:
Pelvic X-rays (PXRs) are essential diagnostic tools used to visualize the
pelvic region and assess pelvic fractures. The rising incidence of pelvic
fractures leads to increased radiologist workload and initial misdiagnoses.
As a result, there is a growing need for automated tools to assist doctors
in pelvic fracture detection. Artificial intelligence has advanced
recently, resulting in several methods for diagnosing PXRs for fractures.
However, concerns regarding annotation accuracy and the limitations of
PXRs due to constrained viewing angles persist. Some fractures are only
visible in 3D computed tomography (CT) images, and it is difficult to
understand their visibility in PXR. This study proposes a method for
using annotations from pelvic CT to label PXRs, focusing on fracture
visibility. Additionally, the impact of labeling PXRs based on visibility
to fracture detection performance in PXR images is examined. First,
all fractures in CT images are annotated using a 3D surface annotation
approach. Next, annotated pseudo PXRs are synthesized from CT
images utilizing digitally reconstructed radiographs (DRRs). The annotated
pseudo PXRs serve as references for accurately labeling fractures
in corresponding PXRs. By training a Resnet-101-based deep convolutional
neural network (DCNN) with the labeled datasets considering
fracture visibility, the proposed method significantly improved fracture
detection performance, achieving an Area Under the Receiver Operating
Characteristic (AUROC) of 0.9114. The AUROC of the conventional
annotation method was 0.8202.
Full Paper (in PDF)
|
Exploring Oversampling Techniques for Fraud
Detection with Imbalanced Classes
Sultan Alharbi, University of Technology Sydney, Australia
Abdulrhman Alorini, University of Technology Sydney, Australia
Khaled Alahmadi, University of Technology Sydney, Australia
Hadeel Alhosaini, University of Technology Sydney, Australia
Yeqian Zhu, University of Technology Sydney, Australia
Xianzhi Wang, University of Technology Sydney, Australia
Abstract:
Each year, credit card fraud has caused significant losses for financial institutions
and individuals worldwide. Financial institutions must detect
credit card fraud to prevent customers from being charged for products
they did not order. Class imbalance has been a standing challenge for
credit card transactions, as the number of fraudulent transactions is significantly
lower than that of non-fraudulent transactions. In this paper,
we comprehensively evaluate five oversampling techniques, namely Synthetic
Minority Oversampling Technique (SMOTE), Adaptive Synthetic
Sampling (ADASYN), Borderline SMOTE, Random Oversampling, and
SMOTE Support Vector Machine (SMOTE SVM), in combination with
seven machine learning techniques (namely XGBoost, Random Forest,
K-Nearest Neighbor, Naive Bayes, Support Vector Machine, LightGBM,
and Convolution Neural Network). Our results show oversampling generally
improves fraud detection performance and SMOTE SVM is the
better oversampling method than other methods under test. Notably,
it achieved an accuracy of 76.47% when used with KNN on the smaller
dataset and 99.93% with CNN on the larger dataset used in our experiments.
Full Paper (in PDF)
|
Generation of Clothing Items with Jamdani Motif
Elements Using Automated Generative Adversarial
Networks
Hujaifa Islam, Samiur Rahman Abir, Md. Sakibur Rahman, Hasan Mahmud, Mohammad Shafiul Alam,
Ahsanullah University of Science and Technology, Bangladesh
Abstract:
Clothing serves as an artistic medium for humans to express their preferences,
thoughts, and cultural heritage, while the application of machine
learning, particularly Generative Adversarial Networks (GANs),
remains largely unexplored in the realm of clothing production and design,
with designers currently relying on their imaginative skills to create
diverse styles. In this article, Conditional Generative Adversarial Networks
(cGAN) are used to suggest an automated approach. Neural style
transfer and cGAN algorithms are employed. to create traditional clothing
with distinctive patterns and a variety of styles. For this study,
the Fashion MNIST and Jamdani Motif Dataset datasets were both employed.
The conditional GAN model was used to produce several styles
of apparel using the MNIST dataset. The Neural Style Transfer model
is then used to combine the created picture with the Jamdani Motif pattern
from the Jamdani Motif dataset. Using Otsu's image segmentation
technique, the foreground, and background of the resulting picture are
separated. Performance scores of this model are as follows: Inception
Score is 1.3573909, Frechet inception distance is 1272.222597, Kernel Inception
Distance is 636200.667, Coverage Metric is 33.79799. We polled
several people on our work output, and the results are detailed in a later
section. Generate Jamdani clothing using single pattern and remove extra
regions using image segmentation.
Full Paper (in PDF)
|
|
