E-Thesis & Research สถาบันเทคโนโลยีไทย-ญี่ปุ่น

Thesis (MIT) 2024

1.	OBJECT DETECTION FOR RETAIL PRODUCT RECOGNITION [แสดงบทคัดย่อ] [ซ่อนบทคัดย่อ]
	ผู้แต่ง : Nakul Pannoy
	The retail sector is a vital driver of economic expansion. The retail industry must embrace technological advancements to augment productivity, streamline operations, and minimize human errors to uphold its crucial economic role. During the COVID-19 pandemic, the purchasing power volumes globally grew from February 2020 to April 2021, and the retail sector gained 35 percent in market capitalization. This growth underscores a compelling research opportunity in the retail industry. Consequently, artificial intelligence (AI) has become a focus of significant interest and has widely adopted technology, which includes computer vision to recognize and detect retail products. This study explores the efficacy of YOLO (You Only Look Once) in retail product recognition. The research compares different YOLO versions and evaluates their ability to identify on-shelf grocery items. In this research, YOLOv8 is applied and divided into five subcategories: YOLOv8n (nano), YOLOv8s (small), YOLOv8m (medium), YOLOv8l (large), and YOLOv8x (extra-large). The models are assessed utilizing Grozi-120 and SKU110K datasets. The evaluation metrics are precision, recall, mAP50, and mAP50-95. The result demonstrates that YOLOv8x yields the best overall performance across both datasets. Remarkably, a mAP50 score of 92.6 percent is achieved on the SKU110K. In conclusion, the outcomes indicate that the YOLOv8 model works well for retail product detection. The model efficacy shows potential for effectiv Full Text : Download!

2.	A MACHINE LEARNING ALGORITHMS FOR DRUG IDENTIFICATION: ENHANCING MEDICATION SAFTEY IN THAILAND [แสดงบทคัดย่อ] [ซ่อนบทคัดย่อ]
	ผู้แต่ง : Sorayus Kittisorayut
	Medication errors are among the leading causes of patient harm within the healthcare system. To address this critical issue, a variety of tools and solutions have emerged to minimize the occurrence of these errors, particularly those stemming from Look-Alike, Sound-Alike (LASA) drugs. Notable among these solutions are automated dispensing systems that utilize barcode scanning and RFID technology to enhance accuracy in medication administration. This study investigates the challenge of drug identification through a machine learning approach. Initially, a dataset comprising commonly used medications in Thailand was manually curated. Subsequently, various versions of the YOLO (You Only Look Once) object detection models were trained and evaluated to assess their efficacy in drug identification. The primary goal of this research is to develop and compare the performance of these different YOLO model variants. Evaluation metrics for this study include precision, recall, F1-score, and mean average precision (mAP). The results demonstrate that the YOLOv8-nano model achieves the highest accuracy in drug identification, recording a mAP score of 99.5%, a recall of 99.7%, and an F1 score of 99.4%, outperforming other versions. In conclusion, this research highlights the significant potential of machine learning as a critical component in efforts to mitigate medication errors in the future. By improving drug identification processes, these advanced technologies can enhance patient saf Full Text : Download!

3.	INTEGRATING EYE-TRACKING INSIGHTS WITH USER FEEDBACK: A UX EVALUATION OF WEBSITE LANDING PAGE [แสดงบทคัดย่อ] [ซ่อนบทคัดย่อ]
	ผู้แต่ง : Wasin Artsawapongtanet
	Eye-tracking technology has become increasingly important in user experience (UX) research due to its ability to capture precise data on how users interact with digital interfaces. While traditional feedback methods like questionnaires provide valuable insights into user perceptions and satisfaction, they are limited by subjective biases. Eye-tracking offers an objective way to measure user attention and engagement. In this paper, we explore the integration of eye-tracking technology with traditional questionnaires to assess user interaction on a redesigned landing page for a university faculty website. By combining these two methods, we aim to uncover the correlation between eye-tracking data and user feedback, particularly in how users process design elements such as layout, colors, and content organization. Using both qualitative and quantitative analysis, the proposed system evaluates user engagement to inform better UX design decisions. The findings demonstrate the feasibility of using eye-tracking as a complementary tool to traditional questionnaires, enhancing the understanding of user behavior and supporting more effective website design strategies. Full Text : Download!

4.	PERFORMANCE ANALYSIS OF IMAGE CLASSIFICATION BETWEEN EDGE AND CLOUD COMPUTING [แสดงบทคัดย่อ] [ซ่อนบทคัดย่อ]
	ผู้แต่ง : Natthasak Vechprasit
	Image classification has become a major application in the AI era for connecting the physical and digital worlds. However it requires intensive graphic processing power. IoT and Edge Computing have become popular approaches for distributing and offloading the workload from the cloud to the edge. Many edge devices are powered by an energy-efficient processor that cant execute intensive workloads but some may be able to. In this thesis we studied the overview of image classification implementation on edge and cloud computing and analyzed the performance to reveal the opportunity to implement a proper system architecture. The edge and cloud computing environments studied in this paper are a smartphone a personal computer and a cloud GPU instance with sample applications to simulate real-world scenarios. The performance is based on the datasets and the processing environment comprising three factors: ML runtime hardware and network. Resulting in six factors: inference time end-to-end execution time (including network delays) accuracy confidence score resource usage and data transfer. Full Text : Download!

5.	DETECTION OF SEIZURES AND THEIR SIMILARITY IN DOGS USING DEEP LEARNING TECHNIQUE [แสดงบทคัดย่อ] [ซ่อนบทคัดย่อ]
	ผู้แต่ง : Thanyasiri Nisaika
	Seizures in dogs are a prevalent and potentially life-threatening neurological issue and early detection is essential for timely treatment and improved outcomes. The proposed research presents a novel deep learning-based method for automatic seizure detection in dogs using video data. This methodology employs a pose estimation technique using the DeepLabCut software and the ResNet50 model which is trained using supervised learning and extracts the (x y) coordinates of the significant anatomical points on the dogs body from the video frames. These keypoint coordinates are input to different types of Recurrent Neural Network (RNN) architectures including SimpleRNN GRU LSTM and Bidirectional LSTM. The RNN models must categorize the dogs activity as either a seizure or a non-seizure activity (such as sitting lying down walking running or jumping). Experimental results show that the GRU model exhibited the best performance with a high Accuracy of up to 90% and a very significant Recall of up to 100% for seizure detection. This indicates the models ability to detect all instances of seizures completely. It performs better than other RNN models that have been tried and also outperforms traditional methods like SVM which only achieved 82.5% accuracy and a recall of 85%. The result creates the potential for home monitoring systems capable of offering real-time seizure detection and fast automatic alerting to owners even where continuous direct monitoring is not feasible. This signific Full Text : Download!

Total 5 Record 1

ABOUT

CONTACT INFO