KMITL

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

Browse

Search Results

Now showing 1 - 10 of 36

Some of the metrics are blocked by your
consent settings
Utilizing deep learning from mobile phone photos for early detection of horizontal strabismus: a screening approach
(2026-12-01)
Chawuthai, Rathachai
;
Sermswan, Anawat
;
Boonnithititikul, Chatree
;
Hokierti, Kiatthida
;
Sermsripong, Wasawat
To develop and validate an artificial intelligence pipeline for binary screening of horizontal strabismus versus orthotropia using smartphone-acquired facial images and geometric landmark analysis. This two-stage system combines Real-Time Detection Transformer (RT-DETR) to localize nine ocular landmarks per eye across three gaze directions (left, center, right), and supervised machine learning classifiers. A feature set of five biometric ratios was derived from coordinates including the canthi, limbi, and corneal light reflexes. The model was trained on facial images from 150 participants (96 with strabismus and 54 controls). To address class imbalance and improve generalizability, Synthetic Minority Oversampling Technique (SMOTE) and 4-fold cross-validation were applied. RT-DETR achieved an intersection over union of 0.62 and a mean center-point error of 6.52 pixels in landmark localization. The Random Forest classifier achieved an accuracy of 0.95, sensitivity of 0.96, specificity of 0.94, positive predictive value of 0.97, and negative predictive value of 0.92. This study demonstrates the feasibility of combining transformer-based landmark detection with geometric ratios for strabismus screening. The framework shows high performance under controlled conditions. While the use of biometric ratios allows for feature-level inspection, further research is required to establish full clinical interpretability and performance in uncontrolled environments.
Some of the metrics are blocked by your
consent settings
Assessing the Effects of Corrupted Parameters in a Large Language Model: A Case Study of LLAMA 3.2 1B
(2026-01-21)
Chawuthai, Rathachai
;
Thongsawaeng, Anon
;
Perdio, John Paul Layug
;
Zaw, Kaung Khant
;
Kraichoke, Phalat
This study explores the effects of parameter corruption in a large language model (LLM) by altering its weights and evaluating performance. Experiments involve corrupting different layers and matrix types, including Self-Attention and Feed-Forward components, with performance assessed using BERT and ROUGE scores. Testing the Llama-3.2-1B-Instruct model was performed on the GLUE-QNLI dataset. Results show that increased corruption leads to greater degradation, with Feed-Forward matrices having a stronger impact especially in the Down matrices. According to the study, later layers are more important for performance than those that come before them. These results shed light on possible future chip implementations of LLM, which may help guide the design of fault-tolerant systems by taking vulnerable parameter placement into account.
Some of the metrics are blocked by your
consent settings
Answer Set Programming for Bus Route Finding: A Case Study in Bangkok, Thailand
(2026-01-01)
Racharak, Teeradaj
;
Mon, Su Yee
;
Chawuthai, Rathachai
Finding user-satisfied bus routes subject to the minimal travel time and the number of route transfers is a challenging problem in transportation networks. In our context, bus routes and bus stations form a directed graph. Thus, a solution to finding the user-satisfied bus routes is a route path between two given stations satisfying a collection of different user's constraints. While this problem could be related to existing research, such as, the study of graph traversal and path finding algorithms. We show that our problem is more suited to formalizing the navigation system in terms of several related answer set programs (ASPs), where each module has its own isolated function, rather than treating like common graph traversal problems such as the well-known Dijkstra's algorithm, which only provides a statically single solution with lowest cost. Our empirical study with a real-world bus network data confirms that our ASPs encoding works in producing a wide range of high-quality solutions, offering more flexibilities between different optimization criteria, compared with the fixed Dijkstra's optimization results. The findings show promise for extending to richer transit representations and optimization objectives beyond the two considered here.
Some of the metrics are blocked by your
consent settings
Novel method for predicting the cracks of oxide scales during high temperature oxidation of metals and alloys by using machine learning
(2025-12-01)
Chawuthai, Rathachai
;
Promchan, Teeratat
;
Rojsanga, Jularak
;
Chandra-ambhorn, Somrerk
;
Nilsonthi, Thanasak
Material degradation is one of the main problems in various high-temperature processes, directly resulting in the failure of the material. Crack and protective oxide film spallation caused either by mechanical stress development in the oxidation process or thermal stress due to a mismatch of the thermal expansions of the formed oxide and alloy are common forms of failure in high-temperature processes. Typically, the Pilling-Bedworth ratio (PBR) is employed to predict crack and spallation of the oxide by determining the volume changes of oxide and alloy because of its simplicity. However, this approach provides poor crack and spallation predictions. Hence, machine learning was adopted in the present work to predict oxide formation and spallation in the temperature range of 600-1,200 °C. The inputs for the present developed model were alloy compositions, oxide formed during oxidation, and oxidation conditions and periods. Furthermore, the predicted results of the present developed machine learning model were compared to those obtained by the PBR method. The present results revealed that the accuracy of the oxide spallation prediction of the present model was better than that of the PBR method. The random forest with 15 estimators was the best machine learning model. Finally, it can be concluded that the machine learning model is essential for accurate material failure prediction.
Some of the metrics are blocked by your
consent settings
Minimizing Model Size of CNN-Based Vehicle Make Recognition for Frontal Vehicle Images
(2025-01-01)
Puisamlee, Wiput
;
Chawuthai, Rathachai
Vehicle Make Model Recognition (VMMR) is commonly used in Intelligent Transportation Systems (ITS), free-flow image-based toll systems, and enforcement systems. These systems must analyze and process vehicle front images for use as evidence. Convolutional Neural Networks (CNN) are widely used for image classification and VMMR problems. Complex model structures and more internal parameters are needed to improve classification accuracy with many classes. Issues included larger models and longer processing times. The goal of this work is to study and create a smaller CNN model that can be used on devices with limited resources, like embedded computers and embedded computer cameras, to figure out what kind of car it is from a front view picture. Real free-flow toll systems were used to train a CNN model that recognized vehicle makes with 99% accuracy. The model is smaller than VGG16, InceptionV3, Yolo11m-cls, and ResNet50 and has over 90% accuracy. It reduced parameters by 69.95% and developed the CTv1 model to achieve an F1 score 2.06% higher than InceptionV3, the best. The model was tested on a Raspberry Pi 3 Model B, processing images in 1 second and using 25 mWh. The compact version of the proposed model also adjusts the Padding and Stride of the Convolutional Layer and reduces the CNN model size using Depth-wise Separable Convolutional and 1 × 1 Convolutional Dimension Reduction (Bottleneck) methods to test vehicle make recognition accuracy, training time, processing time, and model size.
Some of the metrics are blocked by your
consent settings
Integration of Genetic Algorithm with Machine Learning for Properties Prediction
(2025-01-01)
Chawuthai, Rathachai
;
Murathathunyaluk, Siripan
;
Amornratthamrong, Nalin
;
Arunchaipong, Run
;
Anantpinijwatna, Amata
Numerous studies have demonstrated that machine learning (ML) provides more accurate estimations of properties for oxygenated organic derivatives compared to the conventional Quantitative Structure-Property Relationship (QSPR) method. Consequently, ML’s predictive capabilities have been extended to encompass a broader range of properties, including Partition Coefficient, Boiling Point, and Solubility, among others, for oxygenated hydrocarbon derivatives. Algorithms such as Linear Regression, Support Vector Machine, Random Forest, and Gaussian Process are selected through trial-and-error to identify the most suitable approach. The models are trained and validated using experimental data from published literature. Despite the accuracy of these property predictions, they have limited practical utility in industry, where specific property ranges are essential for processes. To address this, Genetic Algorithms (GA) are employed to design chemical compounds that meet industrial requirements. Integrating GA with ML could yield alternative chemical compounds, enhancing overall production processes by increasing economic potential, sustainability, and reducing environmental impact.
Some of the metrics are blocked by your
consent settings
A Comparative Study of Video Segmentation Techniques for Graduate Detection in Thai Graduation Ceremonies
(2025-01-01)
Chungmarisakul, Chanasorn
;
Chawuthai, Rathachai
Graduation ceremonies are significant occasions usually documented on lengthy, difficult-to-navigate videos. To make more satisfaction of the video needs to remove other participants and restore clear areas into short segments focused on particular graduates. By using YOLOv8 for efficient participant segmentation and the Segment Anything Model (SAM2) for accurate tracking, this study expands on earlier research by preparing videos for inpainting. The results establish the foundation for a complete system that combines inpainting, segmentation, and detection to produce polished, customized graduation video clips.
Some of the metrics are blocked by your
consent settings
Modular LLM Architecture with Pluggable Reasoning Heads: A Scalable Approach to Multi-Modal AI Reasoning
(2025-01-01)
Pathak, Anurag
;
Sharma, Dilip Kumar
;
Agrawal, Harshada
;
Chawuthai, Rathachai
;
Petchhan, Jirayu
This paper introduces a modular architecture for Large Language Models (LLMs) that incorporates pluggable, domain-specific reasoning heads to augment the model's capabilities beyond conventional text generation. Central to our approach is an attention-routing controller that intelligently classifies and dispatches user prompts to appropriate reasoning modules - namely symbolic, logical, or graph - based heads-based on the prompt's structure and intent. This design enables hybrid, multi-paradigm reasoning without requiring retraining or finetuning of the base LLM. By decoupling reasoning tasks from general language understanding, our system improves both computational efficiency and interpretability. We demonstrate the architecture using Groq as the base LLM and integrate lightweight engines such as SymPy for symbolic mathematics and custom modules for logical and graph-based reasoning. Experiments across a suite of structured and unstructured prompts show a significant reduction in inference latency and token usage, along with higher accuracy and better explainability. The proposed framework offers a scalable foundation for embedding modular reasoning capabilities into modern LLM-driven applications.
Some of the metrics are blocked by your
consent settings
Detecting AI-Generated Scientific Abstracts Using Galactica and Graph Neural Networks
(2025-01-01)
Pathak, Anurag
;
Sharma, Dilip Kumar
;
Agrawal, Harshada
;
Chawuthai, Rathachai
;
Petchhan, Jirayu
The rise of large language models has introduced new challenges in maintaining research integrity, particularly through the potential proliferation of AI-generated scientific content. This paper presents a novel hybrid framework that combines Galactica-a scientific language model developed by Meta AI-with Graph Neural Networks (GNNs) to detect AI-generated research abstracts. Leveraging the AI-GA dataset comprising 28,662 labeled abstracts, we extract domain-specific semantic embeddings using Galactica and construct a semantic similarity graph to approximate citation-like relationships. A two- layer GCN is then trained to classify each abstract as human- or AI-authored. Experimental results demonstrate that our method outperforms traditional baselines such as TF-IDF, RoBERTa, and perplexity-based detectors, while offering interpretable and scalable detection suitable for editorial screening pipelines.
Some of the metrics are blocked by your
consent settings
Stateless System Performance Prediction and Health Assessment in Cloud Environments: Introducing cSysGuard, an Ensemble Modeling Approach
(2024-01-01)
Chairatana, Nutt
;
Chawuthai, Rathachai
Stateless cloud computing presents remarkable scalability and cost-effectiveness by offering dynamically adjustable resources tailored to fluctuating demands, eliminating the constraints of stateful architectures. However, the challenges presented by dynamic workload are substantial in the context of system health monitoring, frequently leading to service interruptions owing to insufficient resources. It underscores the need for the development of more efficient monitoring systems. Our study introduces cSysGuard, a novel framework designed to enhance monitoring capabilities within cloud environments. The methodology employs an ensemble regression model with a stacking strategy to forecast dynamic performance metrics. The algorithm also leverages a classification model to assess the system's health based on forecasted metrics, effectively identifying potential failures in the future. Under the configuration utilized, our evaluations demonstrated increased predictive performance with cSysGuard in forecasting various metrics compared to traditional models. The results showed an improvement of up to a remarkable 2.28-fold increase, varying significantly based on the specific metric under consideration. In addition, the effectiveness of health assessment was achieved through Decision Trees with hyperparameter tuning, resulting in a macro-averaged F1 score of 89.79%. This research contributes to both the theoretical and practical aspects of server monitoring, presenting a solution that assesses system performance metrics and health to tackle dynamic challenges in cloud infrastructure.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results