DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
LVLMs hallucination on sensor data detection and mitigation
No full textLarge Vision-Language Models (LVLMs) have achieved impressive performance across a range of tasks; however, their reliability in processing sensor data—such as depth and infrared images—remains underexplored. To address this gap, the present study introduces the first benchmark specifically designed to evaluate hallucination phenomena in LVLMs under sensor modalities. Focusing on two common types of sensors—depth cameras and infrared cameras—this study constructs a composite test set comprising over 500 aligned image pairs and develops task scenarios tailored to the unique characteristics of each sensor type. To compensate for the lower resolution and limited information in depth images, bounding boxes are used for target annotation. To ensure controllable model outputs and enable automated evaluation, most benchmark tasks are framed as binary classification problems, requiring the model to select between two candidate answers, while others involve predicting single-digit numerical outputs. Experimental results demonstrate that LVLMs exhibit a substantially higher hallucination rate when handling sensor data compared to conventional RGB images. Nevertheless, the use of lightweight strategies such as Chain-of-Thought (CoT) prompts and Meta prompts significantly reduces hallucinations and improves task performance.Master's degre
Development of smart electronic application
No full textMicrocontrollers are the brains of many IOT devices used today. Making full use of this
technology to simplify and make our lives easier while keeping it affordable. The amount
of effort needed to start growing plants is a huge turnoff for some people wanting to step
into the realm of cultivating home growth plants. This project aims to create an
affordable yet efficient closed monitoring feedback system for growing plants at home. In
this project we integrate both hardware and software to develop a close feedback
monitoring system for users. For our hardware, we use sensors and a microcontroller
called ESP 32. The ESP32 is a powerful, low-cost microcontroller with built-in Wi-Fi
and Bluetooth capabilities. It is widely used in IoT projects for wireless communication
and sensor integration. For software, we will be using Arduino IDE to communicate with
our ESP 32 and Bylink to send notifications to our users. The Arduino IDE is a software
platform used to write, compile, and upload code to Arduino boards. It uses C/C++
programming languages to provide an easy-to-use interface with built-in libraries and
tools, allowing users to interact with microcontrollers.Bachelor's degre
Low dose computated tomography denoising via deep learning
No full textReducing patient radiation exposure while preserving diagnostic precision is a key objective of low-dose computed tomography (CT) imaging. However, these scans often suffer from significant quantum noise, which compromises image quality and hinders accurate diagnosis. To tackle this issue, this study explores cutting-edge deep learning models designed for the denoising of low-dose CT images. We assess the performance of seven state-of-the-art models: RED-CNN, EDCNN, DnCNN, CTformer, MaskedDenoising, Noise2Void, and WGAN-VGG. The evaluation is conducted using the 2016 Mayo Clinic Dataset, a widely acknowledged benchmark for low-dose CT denoising.
The experimental process involved preprocessing the dataset by converting DICOM files into formats compatible with the respective models. Each model was rigorously tested and compared using quantitative metrics to measure denoising effectiveness. Furthermore, qualitative assessments were performed through visual inspections to evaluate the preservation of anatomical details and overall structural integrity in the denoised images.
Key findings include:
– MaskedDenoising and CTformer demonstrated superior performance in suppressing noise while preserving fine details, achieving high PSNR and SSIM scores.
– WGAN-VGG excelled in producing visually appealing outputs by leveraging generative adversarial networks (GANs) combined with VGG-based perceptual loss, ensuring both realism and clinical relevance.
– Noise2Void showed remarkable self-supervised capabilities, effectively denoising images without requiring paired noisy-clean datasets.
– RED-CNN and DnCNN performed well in terms of computational efficiency and generalization across diverse patient data.Bachelor's degre
High efficient organic solar cell
No full textThis report presents the working principle of organic solar cells (OSCs) and ways to
improve its power conversion efficiency (PCE) through simulating it’s operations
with different device structures.
With the aid of simulation software, the OSCs with different layer structures are
designed and their operation are simulated. Collected simulation data are then be
analyzed to observe how each layer affects the overall PCE % of the structure. By
varying the layer thickness, the optimal thickness of the OSC device structure that
produces the best power conversion efficiency can be obtained. With the optimized
layer structure, the OSC achieves the highest PCE of 16.78%.Bachelor's degre
RM2: answer counting queries efficiently under shuffle differential privacy
No full textDifferential privacy (DP) is a leading standard for protecting individual privacy in data collection and analysis. This paper explores the shuffle model of DP, which balances privacy and utility by allowing users to send messages to a trusted shuffler before reaching an untrusted analyzer anonymously. We focus on efficiently implementing the matrix mechanism in shuffle-DP, where efficiency is defined by the number of messages each user sends. Our contributions include a baseline shuffle-DP mechanism that naively adapts the matrix mechanism, followed by an improved mechanism that reduces message complexity while maintaining error levels comparable to central-DP. We demonstrate the versatility of our approach across common query workloads, such as range queries and data cubes, achieving significant improvements in message efficiency. Experimental results confirm that our method outperforms the baseline solution while closely matching the accuracy of central-DP mechanisms.Nanyang Technological UniversityPublished versionThis work has been supported by HKRGC under grants 16205422, 16204223, and 16203924; and by NTU-NAP startup grant 024584-00001
Mastering the medium and the message, riding the platform era
No full textEmerald Hill – The Little Nyonya Story, a 2025 Singaporean drama, has gained a wider cross-border audience by tapping into the current platform-led media environment. The show’s popularity signals a shift in how cultural content is circulated and soft power enhanced. This goes beyond state diplomacy, market design, or fan mobilisation, to include activating the power of today’s media infrastructure and algorithms, through mastering both medium and message.Published versio
Development of a personal activities and events reminder app – part B
No full textDue to the heavy reliance of technology in this modern day, calendar applications
have become an essential for individuals to keep track of their upcoming tasks and
events. Regardless of age and health conditions, people still occasionally had
forgotten to perform a planned action at a future time. There are various Calendar
applications, but none of them address location-based and habitual types of
prospective memory failures.
Therefore, the objective of this project is to design and develop a comprehensive
personal activities and events reminder app, providing users all the features of a
reminder app, by addressing the issues of prospective memory.
This project utilizes the combination of React Native and Expo open-source
frameworks to provide a cross-platform solution. The backend of the app is
supported by Google’s Firebase due to its real-time data capabilities. Implementation
focuses on the development of complex algorithms, including geofencing services,
Google Calendar synchronization and various notification scheduling. The
application offers users an effective solution to manage their commitments.Bachelor's degre
2D to 3D shape morphing dynamic structure
No full textInspired by biological morphogenesis, this study draws on the shape morphing abilities of biopolymers induced by ionic crosslinking and strain mismatch. Stimuli-responsive morphing structures have gained tremendous attention due to their vast potential as alternatives for constructing complex biological tissues and dynamic hydrogel scaffolds. Past research has brought to light a multitude of configurable biomaterials, in particular naturally derived ionic polysaccharides and protein-based gums. However, pioneering experiments have mostly utilised inhomogeneous multi-material precursors such as bilayer hydrogels and composite elastomers, which demands more refined actuation parameters and prolongs fabrication time.
Past studies have also examined a large array of stimuli, particularly in the area of soft robotics and flexible electronics. These techniques have since been cross-experimented on biomaterials, including but not limited to external electric fields, external magnetic fields, film deposition, pneumatics, temperature, humidity, pH and light. Single and combined stimuli techniques induce varying levels of strain mismatch within the structures, morphing 2D precursors into functional 3D structures.
This paper explores the shape morphing abilities of homogeneous sodium alginate using crosslinking agent calcium chloride and CO2 laser-cutting technique to induce strain mismatch. At each of the three stages, an elimination method is elucidated to narrow down and quantify laser-cutting parameters and calcium chloride concentrations. The final stage demonstrates a range of shape morphing and provides key insights between laser-cutting and shape morphing geometric parameters as a function of time. It is hoped that data collected in this study contributes to the understanding and construction of programmable shape morphing biomaterials.Bachelor's degre
See you later: Holding on to life
No full text‘See You Later’ is a 2D digitally hand-drawn animated short film that visualizes the damaging effect of apathy on relationships through a world-ending calamity. More specifically how an individual’s apathy towards life itself causes them to turn to reclusive behaviour, straining the relationships they have with others. Using sunlight as an integral part of the calamity is a deliberate choice for the viewer to more easily imagine themselves inside this alternate universe. Where staying indoors and in the dark, leads to death and outside in daylight, life. This report will document the entire production pipeline from ideation to completion as well as the reasoning behind the decisions made to support the message of the film.Bachelor's degre
A letter red dot
No full textPublic art in Singapore has grown to reflect the nation’s identity, culture, and innovation. However, typography-themed installations remain underrepresented. A Letter Red Dot bridges this gap by reimagining type as a spatial medium — not just for reading, but for experiencing it. This project explores how letters can be functional, poetic, inviting casual interaction and quiet reflection. Inspired by familiar elements of Singapore’s built environment and the cultural memory of “home,” it transforms the A–Z alphabet into inhabitable forms that carry memory and meaning. By merging typography, spatial design, and placemaking, this work restores emotional presence to everyday spaces and encourages deeper engagement with the city’s visual and cultural identity.Bachelor's degre