Thailand Electrical Engineering Journal (TEEJ)

Application of Lightweight Deep Learning Models for Plant Disease Classification via LINE Chatbot

Jantana Panyavaraporn — 2026-01-05

Smart farming is becoming increasingly important in Thailand, especially with advances in artificial intelligence (AI) technology that help reduce reliance on experts and accelerate plant disease diagnosis through leaf images. Therefore, this paper presents a performance comparison of different deep learning models, namely ResNet18, EfficientNet-B0, and MobileNetV3, in classifying cassava diseases using a five-class cassava leaf image dataset and tomato diseases using a ten-class tomato leaf image dataset. Experimental results show that EfficientNet-B0 achieved the highest accuracy and F1-score, followed by MobileNetV3 and ResNet18, respectively. After, EfficientNet-B0 was applied in real-time system, and a prototype was developed by integrating the EfficientNet-B0 model with a LINE chatbot. The system allows users to submit cassava or tomato leaf images and receive automated disease predictions with preliminary treatment recommendations via LINE. The results show that the system performs efficiently and has potential for smart agriculture in the future.

A Study on Impact of Technical Data Communication for Integrating Underwater Drones and Virtual Reality Digital Twin on 5G Network to Enhance Thailand’s Modern Tourism

Sarun Duangsuwan — 2026-01-05

Thailand’s marine tourism industry has transitioned into the digital age. The creation of metaverse tourism environments can further enhance tourist engagement by integrating advanced innovations such as underwater drone technology, digital twin systems, virtual reality (VR), and 5G communication networks. This paper presents the technical impact of data communication in integrating underwater drones with VR technology over a 5G network to advance modern marine tourism. The proposed communication system is designed to enable underwater drones to transmit real-time underwater video and audio to VR headsets via a 5G network. Field experiments were conducted along a 1 km to 5 km coastal route from Phra-Chom beach to Koh-Khai Island, Chumphon province. The technical performance of data communication was evaluated using key parameters, including received signal strength indicator (RSSI), channel bandwidth, path loss, data rate, and video streaming quality. The results demonstrate that real-time streaming from offshore distances of 1 km to 5 km maintained continuous Full HD video quality. The 5G network operating in the 700 MHz frequency band delivered up to 20 MHz of bandwidth. Moreover, transmission distance was found to influence both path loss significantly and received signal strength. Overall, the findings confirm that integrating underwater drones, VR technology, and 5G networks offers a novel and effective approach to developing immersive metaverse tourism experiences, with strong potential to enhance Thailand’s modern marine tourism industry.

An Application of K-Means and Cross-Correlation Techniques for Facial Emotion Recognition

Suchart Yammen — 2026-01-05

This research proposes a high-performance, explainable, and person-specific architecture for Facial Emotion Recognition (FER) that addresses the computational complexity limitations commonly found in deep learning models. The proposed methodology is based on digital signal processing and leverages K-Means clustering to extract template vector features from critical areas of change. A matched filter set under a two-stage structure is then applied for emotion classification. Experiments conducted on the JAFFE dataset using cross-validation for targeted individuals demonstrate that the proposed architecture can perfectly classify all seven basic emotions with an accuracy of 100%. These results highlight the potential of the proposed approach in building highly accurate, lightweight, and user-adaptive emotion recognition systems.

Volume Deep Face: A 3D Face Descriptor for Face Authentication System

Suttipat Srisuk — 2026-01-05

In this paper, we introduce the Volume Deep Face (VDF), a novel face representation proposed for the face authentication system. VDF provides a fast and compact representation of faces using deep learning, enabling one to encode more distinctive features. Using our proposed method, images can be generated to form a 3D VDF representation or a 2D face descriptor (2DFD). The 3D VDF is created from multiple images in the training set, while the 2DFD is generated from a single image during the testing phase. The matching confidence is evaluated using our new volume matching. Our face authentication system is verified with extensive experiments on the XM2VTS database.

A Guidance System for Selecting Bachelor’s Programs in Computer Science Using Neural Network Techniques

Napat Sae-tieo — 2026-01-05

The objective of this study is to develop an artificial neural network system that assists undergraduate computer science students in selecting their courses by utilizing the Multilayer Perceptron (MLP) technique. This program, which we developed with Python and JavaScript, stores data in MongoDB. Graduates from Rajamangala University of Technology Krungthep who graduated between 2012 and 2020 provided the data used in the modelling. It covers their level of education, gender, and academic performance in the three main subjects.

Thailand will experience severe challenges between 2021 and 2023 as a result of a lack of IT personnel, according to CDG's survey. The unemployment rate in the IT sector has been gradually decreasing and is currently at 1.05%, a slight increase from 1.9% in 2022. First-year students interested in computer science and information technology programs at Rajamangala University of Technology Krungthep, however, are still unsure of how to select a course of study that aligns with their interests and professional objectives.

The MLP1 model with one hidden layer, sixteen neurons, regularization at 0.01 and learning rate at 0.005 produced the best results after varying the number of hidden layers, neurons, regularization, learning rate, batch size, and epochs. The accuracy of this model was the highest at 0.66. The results of the study demonstrate that a simple model structure can work effectively and produce favourable results. This system can assist students in making future course selections.

Development of a Water Temperature Control System for Nano Aquariums Using Fuzzy Logic Principles via a Web Application

Atirarj Suksawad — 2026-01-05

The design and implementation of an intelligent temperature control system for small aquariums utilizing fuzzy logic control combined with an Internet of Things platform is suggested by this study. Utilizing an ESP32 as its central processing unit, the system interfaces with accurate pH and temperature sensors to track environmental conditions. The system uses a fuzzy logic method to dynamically modify a thermal control module that includes a cooling fan, heatsink, and Peltier device. An OLED screen shows real-time data, which is then sent to a specially created web application so that users can remotely monitor and control the aquarium online. The system's capacity to accurately maintain water temperature within the ideal range, adapt to environmental changes, and guarantee data dependability through both local and online interfaces is demonstrated by the experimental findings. The suggested approach lowers long-term operating costs and improves aquarium maintenance efficiency, making it a workable option for intelligent aquatic ecosystems.

Comparative Analysis of Regression Models: A Case Study of KNN Regression vs. Multiple Linear Regression

Kunnika Jakor — 2026-01-05

This research paper presents the application of the K-Nearest Neighbors (KNN) algorithm as a regression model for numerical prediction. We propose for KNN regression as a viable alternative for scenarios characterized by non-linear or ambiguously linear data relationships, where conventional linear regression models frequently underperform. Our experimental findings concentrate on evaluating the efficiency of KNN regression in comparison to established models such as multiple linear regression across five datasets. This illustrates the capability of KNN regression to achieve more accurate numerical predictions. In addition, we explore the effects of distance metrics, the inverse distance weighting (IDW) method for neighbor weighting, and K-value selection (number of neighbors) in our in-depth parameter tuning for KNN regression. The results suggest that KNN regression is an efficient and compelling alternative regression model for numerical prediction, particularly when dealing with complicated data and ambiguous linear correlations. Thus relieves the need for more complex models like artificial neural networks.