Papers by Md Mahbub E Noor
Due to the presence of distortion, most of the single-channel frequency-domain speech enhancement... more Due to the presence of distortion, most of the single-channel frequency-domain speech enhancement (SE) approaches are still challenging for downstream automatic speech recognition (ASR) tasks, even with satisfactory improvements in enhancing speech quality and intelligibility. Recently, transformer-based models have shown better performance in speech processing tasks. Therefore, we intend to explore a transformer-based SE model, which is fine-tuned through a two-stage training scheme. Pre-training is performed using a feature-level optimization criterion through SE loss, and then a pre-trained end-to-end ASR model is used to fine-tune the SE model using an ASRoriented optimization criterion through SE and ASR losses. We evaluate the proposed approach on low-resourced Bengali language, which has not received as much attention as resource-rich English or Mandarin languages in both SE and ASR fields. Experimental results show that it can improve the performance of SE and ASR under severe unseen noisy conditions and its performance is reasonably good compared with other state-of-the-art SE methods.
Diabetic retinopathy (DR) is a severe global problem that affects millions of people worldwide an... more Diabetic retinopathy (DR) is a severe global problem that affects millions of people worldwide and gets worse over time. If left untreated, DR can lead to blindness. Early and precise DR identification is necessary to address this developing challenge. Traditional approaches include applying machine learning or deep learning algorithms directly on the dataset or the preprocessed dataset, which has shown very good results recently. Very few focused on combining machine learning and deep learning-based algorithms. Extracting a good set of features is very crucial to getting higher performance from any machine learning-based or deep learning-based algorithm. This work presents a new method for DR detection by fusing a convolution neural network-based feature extraction method before feeding the data to a stacking ensemble learner, which uses several machine learning algorithms to make it more robust. Predictions from several classifiers, including decision trees, random forests, support vector machines, logistic regression, and others, have been used in previous studies of DR. In our work, we used these classifiers for our hybrid model. First, retinal image features are extracted using InceptionV3. Then, several fine-tuned machine learning-based classifiers have been used. Finally, all the classifier models are stacked together to create an ensemble model. Our hybrid approach showed promising performance in classifying binary (98.64%) and multiclass (94.95%) on the APTOS 2019 Blindness Detection dataset. This finding proves that our
Giving appropriate treatment and assistance to persons afflicted by autism spectrum disorder (ASD... more Giving appropriate treatment and assistance to persons afflicted by autism spectrum disorder (ASD) requires early identification. Previous research indicates that early facial expressions can be a valuable tool for identifying ASD. As people's facial expressions vary from ethnicity to ethnicity, so it is essential to choose a specific ethnicity for building a more accurate model of that ethnicity. The purpose of this study is to create and evaluate how well different deep learning models identify ASD from facial expressions in the Bangladeshi population. Because of this study's primary objective is to examine the facial expressions of toddlers and teenagers from Bangladesh, we produced two datasets with 1500 photos for each of them. We used a variety of cutting-edge deep learning models, such as EfficientNetV2L, EfficientNetB0, EfficientNetB7, MobileNetV2, ResNet50, ResNet101, DenseNet201, VGG16, and VGG19 for this study. Our investigation provided excellent findings, with the EfficientNetB0 model discriminating ASD from non-ASD people with an accuracy of 99% for toddlers and 94% for teenagers. Other models also showed excellent precision, with an average accuracy of 96% for toddlers and 82.5% for teenagers. The results highlight the effectiveness of transfer learning in the Bangladeshi context for detecting ASD using facial expression analysis, laying the groundwork for future studies and possible practical applications.
Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that affects repetitive behavior... more Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that affects repetitive behaviors, social skills, verbal and nonverbal communication, and the ability to acquire new knowledge. Typically manifesting in early childhood, between 6 months and 5 years, the symptoms of ASD can evolve over time. Consequently, diagnosing ASD can occur at various life stages, including childhood, adolescence, and adulthood. The field of medical diagnostics research has increasingly explored the application of machine learning techniques, particularly deep learning algorithms, for predicting and analyzing ASD in children and toddlers. This study aims to compare the effectiveness of several deep learning algorithms-namely Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), Recurrent Neural Networks (RNN), Deep Neural Networks (DNN), and Long Short Term Memory Networks (LSTM)-in diagnosing ASD. By experimenting with two distinct datasets, one for toddlers and one for children, we assessed the performance of these algorithms. Our results indicated that CNN achieved 100% accuracy on the children dataset, while ANN attained 100% accuracy on the toddler dataset. This comparative analysis provides insights into the most suitable deep learning algorithms for Autism Spectrum Disorder prediction in different age groups.
Evaluation of OpenStack (Havana release) and CloudStack(4.3 release) Open Source Cloud Solutions
The phenomenon “cloud computing” started from 2006-07 but it came into existence through a very s... more The phenomenon “cloud computing” started from 2006-07 but it came into existence through a very smooth historical events. From the concept of multiprogramming then virtualization, grid computing, SaaS, utility computing and then finally cloud computing came to today’s IT world. Though, the tools used to build cloud are increasing day by day, so it makes a confusion and difficulty to select a suitable tool for the cloud consumers and enterprises to build their cloud for business purpose due to holding various features. Due to facing this problem, it is now inevitable to differentiate these tools. For this reason, this study is about to build a basic infrastructure of two separate clouds using two open source IaaS (Infrastructure as a Service) cloud tools i.e., OpenStack and CloudStack to check how these two differ from each other. This study will also show their evaluation and future scopes.
Low Power Devnagari Unicode Checker Design Using CGVS Approach
In this paper we have introduced a new approach called Clock Gating and Voltage
Scaling (CGVS), ... more In this paper we have introduced a new approach called Clock Gating and Voltage
Scaling (CGVS), which is the combination of two existing techniques i.e. Clock gating and Voltage
Scaling. Our aim is to design a low power Devnagari Unicode Checker (DUC) using CGVS
technique. This design is implemented on Kintex-7 FPGA families, XC7K70T device, -3 speed
grade and FBG676 package. From our analysis, it is observed that, with the use of clock gated
technique in our target circuit and with the scaling of voltage from 1.0V to 0.1V, we are achieving
clock power reduction of 98.98% on 10GHz and 1THz operating frequencies. Under same voltage
scaling scheme, there is 6.66%, 10.38%, 10.64% and 10.62% less reduction in IO power, when the
target circuit is operating on 1GHz, 10GHz, 100GHz and 1THz operating frequencies.
Frequency Scaling Based Green Mobile Battery Charge Controller Sensor Design on FPGA
Frequency Scaling based energy efficient MBCCS is implemented in this paper. In this
design, if ... more Frequency Scaling based energy efficient MBCCS is implemented in this paper. In this
design, if battery's voltage or current is less than threshold then battery will continue charging.
Whereas, if voltage or current is more than threshold then it will ring Overcharge Alarm. With
Frequency Scaling, we reduce frequency from 1THz to 25GHz, where 125GHz, 625GHz are
intermediate frequency value. There is 97.50%, 100%, 97.54%, 97.48%, 20%, 96.48 % reduction in
clock power(CP), logic power(LP), signal power(SP), IOs power(IOP), leakage power and total
power dissipation respectively when we scale down frequency from 1 THz to 25GHz.
In this paper we are presenting result of simulation based energy efficient bi-directional visito... more In this paper we are presenting result of simulation based energy efficient bi-directional visitor counting machine (VCM) on FPGA (Field Programmable Gate Array). In this work, we have used Xilinx software. We have used different IOs standards that include HSTL_I, HSTL_II, HSTL_I_18, HSTL_II_18, LVCMOS12, LVCMOS15, LVCMOS18, LVCMOS25, and LVCMOS33. For these IOs standard we have collected the total energy dissipation for this bi-directional VCM on FPGA and compared them. It is observed that at 5GHz frequency HSTL_II is the lowest energy dissipation for this Bi-directional Visitor Counting Machine (VCM) on FPGA. FPGA is more effective than using any microcontroller in perspective of energy efficiency.
In this work, we are using Voltage Scaling as energy efficient technique to make energy
efficien... more In this work, we are using Voltage Scaling as energy efficient technique to make energy
efficient MBCCS. In MBCCS, whenever voltage or current of battery will under the threshold level
the battery will continuing the charge if more than voltage or current threshold then battery will
invoke Ring Overcharge Alarm. In Voltage Scaling, we scale down supply voltage from 1V to
0.1V, where 0.9V, 0.8V, 0.7V, 0.6V, 0.5V, 0.4V, 0.3V and 0.2V are intermediate supply voltage
value. There is 92.72%, 87.50%, 98.77% and 68.24% reduction in Clock Power, Logic Power,
Signal Power and IO Power on 1 THz device operating frequency when voltage is scale down from
1V to 0.1V with step size of 0.1V. 
Uploads
Papers by Md Mahbub E Noor
Scaling (CGVS), which is the combination of two existing techniques i.e. Clock gating and Voltage
Scaling. Our aim is to design a low power Devnagari Unicode Checker (DUC) using CGVS
technique. This design is implemented on Kintex-7 FPGA families, XC7K70T device, -3 speed
grade and FBG676 package. From our analysis, it is observed that, with the use of clock gated
technique in our target circuit and with the scaling of voltage from 1.0V to 0.1V, we are achieving
clock power reduction of 98.98% on 10GHz and 1THz operating frequencies. Under same voltage
scaling scheme, there is 6.66%, 10.38%, 10.64% and 10.62% less reduction in IO power, when the
target circuit is operating on 1GHz, 10GHz, 100GHz and 1THz operating frequencies.
design, if battery's voltage or current is less than threshold then battery will continue charging.
Whereas, if voltage or current is more than threshold then it will ring Overcharge Alarm. With
Frequency Scaling, we reduce frequency from 1THz to 25GHz, where 125GHz, 625GHz are
intermediate frequency value. There is 97.50%, 100%, 97.54%, 97.48%, 20%, 96.48 % reduction in
clock power(CP), logic power(LP), signal power(SP), IOs power(IOP), leakage power and total
power dissipation respectively when we scale down frequency from 1 THz to 25GHz.
efficient MBCCS. In MBCCS, whenever voltage or current of battery will under the threshold level
the battery will continuing the charge if more than voltage or current threshold then battery will
invoke Ring Overcharge Alarm. In Voltage Scaling, we scale down supply voltage from 1V to
0.1V, where 0.9V, 0.8V, 0.7V, 0.6V, 0.5V, 0.4V, 0.3V and 0.2V are intermediate supply voltage
value. There is 92.72%, 87.50%, 98.77% and 68.24% reduction in Clock Power, Logic Power,
Signal Power and IO Power on 1 THz device operating frequency when voltage is scale down from
1V to 0.1V with step size of 0.1V.