KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Search Results

Now showing 1 - 10 of 11

Some of the metrics are blocked by your
consent settings
Mathematical Benchmarking of Convolutional Neural Networks for Thai Dialect Recognition: A Spectrogram Texture Classification Approach
(2026-03-01)
Visutsak, Porawat
;
Ongrungruaeng, Duongduen
;
Wiriya, Surapong
;
Ryu, Keun Ho
This study rigorously evaluates 13 Convolutional Neural Network (CNN) architectures for Thai dialect recognition. By treating Automatic Speech Recognition (ASR) as a computer vision texture classification task, we processed an extensive 840-h dataset from the Spoken Language Systems, Chulalongkorn University (SLSCU) corpus. Raw audio from four major dialects—Central, Northern (Khummuang), Northeastern (Korat), and Southern (Pat-tani)—was transformed into 2D Mel-spectrograms using the Short-Time Fourier Transform (STFT). We analyzed a diverse range of architectures, including the VGG, Inception, ResNet, DenseNet, and MobileNet families, to establish the optimal trade-off between mathematical complexity and spectral feature extraction. Our experimental results identify NASNet-Mobile as the most effective model, achieving a macro-average F1-score of 0.9425. The analysis suggests that NASNet’s search-optimized cell structure is uniquely capable of capturing the multiscale texture of phonetic formants. In contrast, we observed a catastrophic mode collapse in VGG16 (32.97% accuracy), likely due to excessive parameter bloat, while Xception and MobileNetV2 maintained robust generalization. Confusion matrix analysis reveals high acoustic distinctiveness for Southern Thai (96.7% recall), whereas Northern Thai exhibits significant spectral overlap with Central Thai. These results support the hypothesis that CNNs interpret spectrograms as textures rather than discrete objects, positioning NASNet-Mobile as a high-performance, low-latency baseline for edge-device deployment in resource-constrained environments.
Some of the metrics are blocked by your
consent settings
Novel Adaptive Intelligent Control System Design
(2025-08-01)
Duanyai, Worrawat
;
Song, Weon Keun
;
Ka, Min Ho
;
Lee, Dong Wook
;
Dissanayaka, Supun
A novel adaptive intelligent control system (AICS) with learning-while-controlling capability is developed for a highly nonlinear single-input single-output plant by redesigning the conventional model reference adaptive control (MRAC) framework, originally based on first-order Lyapunov stability, and employing customized neural networks. The AICS is designed with a simple structure, consisting of two main subsystems: a meta-learning-triggered mechanism-based physics-informed neural network (MLTM-PINN) for plant identification and a self-tuning neural network controller (STNNC). This structure, featuring the triggered mechanism, facilitates a balance between high controllability and control efficiency. The MLTM-PINN incorporates the following: (I) a single self-supervised physics-informed neural network (PINN) without the need for labelled data, enabling online learning in control; (II) a meta-learning-triggered mechanism to ensure consistent control performance; (III) transfer learning combined with meta-learning for finely tailored initialization and quick adaptation to input changes. To resolve the conflict between streamlining the AICS’s structure and enhancing its controllability, the STNNC functionally integrates the nonlinear controller and adaptation laws from the MRAC system. Three STNNC design scenarios are tested with transfer learning and/or hyperparameter optimization (HPO) using a Gaussian process tailored for Bayesian optimization (GP-BO): (scenario 1) applying transfer learning in the absence of the HPO; (scenario 2) optimizing a learning rate in combination with transfer learning; and (scenario 3) optimizing both a learning rate and the number of neurons in hidden layers without applying transfer learning. Unlike scenario 1, no quick adaptation effect in the MLTM-PINN is observed in the other scenarios, as these struggle with the issue of dynamic input evolution due to the HPO-based STNNC design. Scenario 2 demonstrates the best synergy in controllability (best control response) and efficiency (minimal activation frequency of meta-learning and fewer trials for the HPO) in control.
Some of the metrics are blocked by your
consent settings
Optimizing Product Quality Prediction in Smart Manufacturing Through Parameter Transfer Learning: A Case Study in Hard Disk Drive Manufacturing
(2025-04-01)
Kaitwanidvilai, Somyot
;
Sittisombut, Chaiwat
;
Huang, Yu
;
Bom, Sthitie
In recent years, the semiconductor industry has embraced advanced artificial intelligence (AI) techniques to facilitate intelligent manufacturing throughout their organizations, with particular emphasis on virtual metrology (VM) systems. Nonetheless, the practical application of data-driven virtual metrology for product quality inspection encounters notable hurdles, such as annotating inspections in highly dynamic industrial environments. This leads to complexities and significant expenses in data acquisition and VM model training. To address the challenges, we delved into transfer learning (TL). TL offers a valuable avenue for knowledge sharing and scaling AI models across various processes and factories. At the same time, research on transfer learning in VM systems remains limited. We propose a novel parameter transfer learning (PTL) architecture for VM systems and examine its application in industrial process automation. We implemented cross-factory and cross-recipe transfer learning to enhance VM performance and offer practical advice on adapting TL to meet individual needs and use cases. By leveraging extensive data from Seagate wafer factories, known for their large-scale and high-dimensional nature, we achieved significant PTL performance improvements across multiple performance metrics, with the true positive rate (TPR) increasing by 29% and false positive rate (FPR) decreasing by 43% in the cross-factory study. In contrast, in the cross-recipe study, TPR increased by 27.3% and FPR decreased by 6.5%. With our proposed PTL architecture and its performance achievements, insufficient data from the new manufacturing sites, new production lines and new products are addressed with shorter VM model training time and smaller computational power with strong final quality prediction confidence.
Some of the metrics are blocked by your
consent settings
Cross-Domain Robust Liveness Detection: A Transfer Learning Approach for Combating Sophisticated Presentation Attacks in Mobile Authentication
(2025-01-01)
Kerdpramote, Phuvis
;
Poomdeesittinon, Akeanant
;
Jamsri, Rapeeploy
;
Krueyos, Phanuwit
;
Muangkan, Ronnakorn
As biometric authentication systems become ubiquitous in Southeast Asia's digital economy, sophisticated presentation attacks using deepfakes, high-resolution displays, and 3D masks pose critical security threats. This paper presents a comprehensive cross-domain liveness detection framework that addresses the generalization challenges plaguing current systems. Our approach leverages MobileNetV2-based transfer learning with a novel two-phase training strategy, achieving superior cross-domain performance while maintaining computational efficiency for mobile deployment. We introduce domain-aware augmentation techniques and evaluate our system across multiple benchmark datasets including NUAA and a locally-collected Thai demographic dataset. Experimental results demonstrate 84.35% accuracy on NUAA and 78.62% cross-domain accuracy, with significant improvements in Attack Presentation Classification Error Rate (APCER) reduction from 28.7% to 15.4% compared to baseline methods. The system successfully detects emerging attack vectors including deepfake videos and tablet-based spoofing attempts. We provide comprehensive analysis of deployment challenges in resource-constrained environments demonstrating practical applicability for Thailand's mobile banking and digital identity verification ecosystem.
Some of the metrics are blocked by your
consent settings
Meta-Learning-Based Physics-Informed Neural Network: Numerical Simulations of Initial Value Problems of Nonlinear Dynamical Systems without Labeled Data and Correlation Analyses
(2024-06-01)
Duanyai, Worrawat
;
Song, Weon Keun
;
Chitthamlerd, Thanadol
;
Kumar, Girish
There are several main challenges in solving nonlinear differential equations with artificial neural networks (ANNs), such as a nonlinear system’s sensitivity to its initial values, discretization, and strategies for incorporating physics-based information into ANNs. As for the first issue, this paper addresses the initial value problems of nonlinear dynamical systems (a Duffing oscillator and a Burger’s equation), which cause large global truncation errors in sub-domains with a significant reduction in the influence of initial constraints, using meta-learning-based physics-informed neural networks (MPINNs). The MPINNs with dual learners outperform physics-informed neural networks with a single learner (no fine reinitialization capability). As a result, the former approach improves solution convergence by 98.83% in the sub-time domain (III) of a Duffing oscillator, and by 85.89% at t = 45 in a Burger’s equation problem, compared to the latter one. Model accuracy is highly dependent on the adaptability of the initial parameters in the first hidden layers of the meta-models. From correlation analyses, it is obvious that the parameters become less (the Duffing oscillator) or more (the Burger’s equation) correlated during fine reinitialization, as the update manner differs or is similar to the one used in pre-initialization. In the first example, the MPINN achieves both the mitigation of model sensitivity to its output and the improvement of model accuracy. Conversely, the second example shows that the proposed approach is not enough to solve both issues simultaneously, as increased model sensitivity to its output leads to higher model accuracy. The application of transfer learning reduces the number of iterative pre-meta-trainings.
Some of the metrics are blocked by your
consent settings
Event-triggered model reference adaptive control system design for SISO plants using meta-learning-based physics-informed neural networks without labeled data and transfer learning
(2024-04-01)
Duanyai, Worrawat
;
Song, Weon Keun
;
Konghuayrob, Poom
;
Parnichkun, Manukid
This paper examines the controllability of a novel Lyapunov-based model reference adaptive control (MRAC) system designed with a meta-learning-based physics-informed neural network (MLPINN) for linear and nonlinear single-input and single-output (SISO) plants without labeled data (MLPINN-MRAC system). It is devised with the benefits of several techniques: the integration of the identification process in a training mode into an online control mode (straightforward design); no labeled data generation for online identification by a physics-informed neural network; the prevention of degradation in tracking performance by meta-learning, as the system triggers a meta-learning process only when an error threshold detects the deterioration (high efficiency and the reduction of computation cost); quick adaptation to new inputs and an updated control input for each sub-time span by transfer learning. It is worth noting that the frequency of the meta-learning event detection significantly affects the step response stability of the nonlinear plant. To achieve a better quality of the stabilization, more frequent event detection is necessary for both beginning and end intervals in control. Sixteen triggering events are enough to shape the acceptable step response of the nonlinear plant, and 44 triggering events achieve the plant's desired step response with its minor modeling error. While, as for the linear plant, a single triggering event is sufficient to attain its tolerable step response and modeling error, implying that intensive event detection is not critical in the identification. It is obvious that the MLPINN-MRAC system functions well and is more beneficial and efficient for the nonlinear plant.
Some of the metrics are blocked by your
consent settings
Multimodal Biometrics Recognition Using a Deep Convolutional Neural Network with Transfer Learning in Surveillance Videos
(2022-07-01)
Aung, Hsu Mon Lei
;
Pluempitiwiriyawej, Charnchai
;
Hamamoto, Kazuhiko
;
Wangsiripitak, Somkiat
Biometric recognition is a critical task in security control systems. Although the face has long been widely accepted as a practical biometric for human recognition, it can be easily stolen and imitated. Moreover, in video surveillance, it is a challenge to obtain reliable facial information from an image taken at a long distance with a low-resolution camera. Gait, on the other hand, has been recently used for human recognition because gait is not easy to replicate, and reliable information can be obtained from a low-resolution camera at a long distance. However, the gait biometric alone still has constraints due to its intrinsic factors. In this paper, we propose a multimodal biometrics system by combining information from both the face and gait. Our proposed system uses a deep convolutional neural network with transfer learning. Our proposed network model learns discriminative spatiotemporal features from gait and facial features from face images. The two extracted features are fused into a common feature space at the feature level. This study conducted experiments on the publicly available CASIA-B gait and Extended Yale-B databases and a dataset of walking videos of 25 users. The proposed model achieves a 97.3 percent classification accuracy with an F1 score of 0.97and an equal error rate (EER) of 0.004.
Some of the metrics are blocked by your
consent settings
Traffic Light and Crosswalk Detection and Localization Using Vehicular Camera
(2022-01-01)
Wangsiripitak, Somkiat
;
Hano, Keisuke
;
Kuchii, Shigeru
An improved convolutional neural network model for traffic light and crosswalk detection and localization using visual information from a vehicular camera is proposed. Yolov4 darknet and its pretrained model are used in transfer learning using our datasets of traffic lights and crosswalks; the trained model is supposed to be used for red-light running detection of the preceding vehicle. Experimental results, compared to the result of the pretrained model learned only from the Microsoft COCO dataset, showed an improved performance of traffic light detection on our test images which were taken under various lighting conditions and interferences; 36.91% higher recall and 39.21% less false positive rate. The crosswalk, which is incapable of detection in the COCO model, could be detected with 93.37% recall and 7.74% false-positive rate.
Some of the metrics are blocked by your
consent settings
Automatic Sleep Data Scoring by Artificial Intelligence: A Pilot Study in Thai Population
(2021-04-01)
Jansri, Ukkrit
;
Chirakalwasan, Naricha
;
Chaitusaney, Busarakum
;
Busayakanon, Supasuta
;
Khongjui, Thamonwan
Sleep apnea, a sleep-disordered breathing (SDB), is defined as repeatedly intermittent cessation of breathing during sleep. It causes various life-threatening diseases. The American Academy of Sleep Medicine (AASM) releases the manual for sleep data scoring. Patients with SDB are prescribed to be monitored at the sleep clinic where several physiological data are recorded, called polysomnogram (PSG). The massive PSG data must be scored by the well-trained expert before being diagnosed by the physician. Our research is to use the Artificial Intelligence (AI) in sleep data scoring, particularly in respiratory events detection. Three ready-made Convolution Neural Networks (CNN); AlexNet, ResNet-50, and VGG-16, with transfer learning were applied to classify 5 overnight PSG data from Chulalongkorn hospital. Our preliminary results showed that all networks provide higher classification result in European Data Format (EDF) than in the text (ASCII) formats (71% vs 54%). The ResNet-50 model structure performed better than the other two networks on both data formats. As expected, the visualized (EDF) data is better than the unconditioned (ASCII) data. Our future development is modifying learning model to increase the scoring performance from more recruited PSG data.
Some of the metrics are blocked by your
consent settings
Municipal solid waste segregation with CNN
(2019-07-01)
Srinilta, Chutimet
;
Kanharattanachai, Sivakorn
Pollution from municipal solid waste has been a problem in Thailand for a long time. People generate waste in every minute. Ineffective waste segregation does increase difficulties in solid waste management. The Pollution Control Department of Thailand provides segregation guideline for municipal solid waste. Household wastes should be separated into four types-general waste, compostable waste, recyclable waste and hazardous waste. This paper explored performance of CNN-based waste-type classifiers (VGG-16, ResNet-50, MobileNet V2 and DenseNet-121) in classifying waste types of 9,200 municipal solid waste images. Waste type can be identified directly from waste-type classifier or derived from waste-item class. Derived classifiers outperformed their corresponding direct classifiers in the experiment. The highest waste-type classification accuracy was 94.86% from the derived ResNet-50 classifier.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results