DSpace collection: 會議論文

MGGA: Make GeM Great Again via Regularization Branch to Mitigate Channel Vanishing in Visual Place Recognition

title: MGGA: Make GeM Great Again via Regularization Branch to Mitigate Channel Vanishing in Visual Place Recognition abstract: Deep-learning-based methods have achieved significant success in the Visual Place Recognition (VPR) task, which is important for autonomous driving and robotics systems. Recent advancements primarily focus on the sophisticated feature aggregation module. This paper argues for a shift in emphasis toward the backbone features. Through an in-depth analysis of GeM, one of the simplest pooling aggregator based VPR method, we identify a prevalent issue, termed ’Channel vanishing’. The issue manifests as a substantial proportion of channels in both the final GeM descriptor and the backbone output local features turning zero-valued and inactive during training, thereby drastically diminishing the representational capacity of the model and undermining its VPR performance. In order to solve this problem, we propose a regularization branch with a fully connected layer for the GeM pipeline. This branch successfully mitigates Channel vanishing and further enriches the diversity and representation of the backbone output features. During inference, our streamlined model, using only the GeM aggregator, achieves state-of-the-art performance among backbones that are not transformerbased. Notably, when utilizing the DINOv2-B backbone, our method derives 99.1% recall@1 and 100% recall@5 VPR scores on the Tokyo24/7 dataset. This result suggests that strengthening backbone features can substantially narrow the gap between simple GeM pooling and more complex aggregators; assessing how broadly this observation transfers to other aggregators is an interesting direction.

Deep Learning-Based Identification of Rab Proteins: A Convolutional Neural Network Approach with Evolutionary Information Integration

title: Deep Learning-Based Identification of Rab Proteins: A Convolutional Neural Network Approach with Evolutionary Information Integration abstract: Rab proteins play a crucial role in membrane trafficking and are implicated in various human diseases. Accurate identification of Rab proteins within membrane proteins is of utmost importance for comprehending these diseases and establishing effective drug targets. In this study, we applied a two-dimensional convolutional neural network (CNN) integrated with evolutionary information to discern and identify Rab proteins present within general proteins. Our CNN model exhibited notable performance, achieving a sensitivity of 93.3%, specificity of 98%, accuracy of 96.9%, and a Matthews correlation coefficient (MCC) of 0.91 when tested on an independent dataset. In comparison to previously published methodologies, our approach displayed a substantial 25% improvement in the identification of Rab GTPases. These findings underscore the potential of deep learning techniques for accurately discerning Rab proteins and lay the groundwork for future investigations employing deep learning in the field of bioinformatics.

Enhancing Segmentation Performance for Cellular and Subcellular Structures in Micrographs: Leveraging ROI, Neighbor Extraction, and Size Constraints

title: Enhancing Segmentation Performance for Cellular and Subcellular Structures in Micrographs: Leveraging ROI, Neighbor Extraction, and Size Constraints

Optimal YOLO-based Model for Fabric Anomaly Detection

title: Optimal YOLO-based Model for Fabric Anomaly Detection abstract: Fabric anomaly detection is a crucial application in the industry. This study identifies the optimal YOLO (You Only Look Once) algorithm from a selection of YOLO versions for detecting fabric anomalies, including defect identification and region localization. Recent YOLO models, including YOLOv5, YOLOv7, YOLOv8, and YOLOv9, are evaluated with batch sizes of 4, 8, and 16. Additionally, computation times for detection are compared. The dataset is generated from numerous images extracted from a fabric video, with test images categorized as normal, line defect, or hole defect. Experimental results show that YOLOv9 batch 4 achieves the highest F1-score for defect detection, while YOLOv8 batch 16 offers a balance of optimal mAP and reduced training time. Larger batch sizes consistently enhance training efficiency across all models. Further experiments can extend this approach to other fabric datasets to detect various types of defects.

Funding Agencies Identification Using Pre-trained Large Language Models

title: Funding Agencies Identification Using Pre-trained Large Language Models

利用新聞與社群貼文標題的情緒分析用以預測個股股價走勢 – 以台積電為例

title: 利用新聞與社群貼文標題的情緒分析用以預測個股股價走勢 – 以台積電為例 abstract: 本研究探討中文大型語言模型（LLMs）於金融情緒分析在台灣股市應用之可行性，特別聚焦於台積電股價預測。透過蒐集2023年8月至2025年3月間之PTT股市版與鉅亨網關於台積電的標題資料，建立三種以BERT為基礎之情緒分析模型（EC_BERT、Dict_BERT、PN_BERT）。這三種模型分別採用Twitter財經標題、NTUSD中文情緒詞典與本研究自建資料進行訓練。實驗結果顯示，EC_BERT在新聞標題分析上表現最佳，準確率最高達57.14%；而Dict_BERT則於分析PTT貼文情緒用以預測股價時具備穩定預測力。相較之下，PN_BERT因訓練資料量與泛化能力(Generalization)限制，準確率略遜。研究亦發現社群媒體之情緒信號較易與股價變動產生關聯，惟新聞報導提供的穩定性與權威性亦具補充價值。整體而言，本研究驗證了LLMs於中文股市情緒分析之潛力，並強調模型微調訓練資料與應用場景是否相符對預測效能之重要性。

An Improved Data Smooth Method with its Application on Aerial Calligraphy

title: An Improved Data Smooth Method with its Application on Aerial Calligraphy

A Cross-Attention Enhanced Approach for Hybrid Text Summarization

title: A Cross-Attention Enhanced Approach for Hybrid Text Summarization

A Novel Time-Awareness Recommendation System

title: A Novel Time-Awareness Recommendation System

An Efficient Cluster-Based Continual Learning with Gradient Episodic Cache Memory

title: An Efficient Cluster-Based Continual Learning with Gradient Episodic Cache Memory

A Reinforcement Learning-Based Mobile Charging Agent for Wireless Rechargeable Sensor Networks

title: A Reinforcement Learning-Based Mobile Charging Agent for Wireless Rechargeable Sensor Networks abstract: Wireless Rechargeable Sensor Networks (WRSNs) are crucial for many applications, including environmental monitoring, healthcare, and smart cities. However, optimizing the charging schedule of a mobile charger to maximize network coverage while minimizing charging latency and energy consumption remains a challenge. In this paper, we design a reinforcement learning-based mobile charging agent (RLMCA), which integrates dynamic window search (DWS), Q-learning, and SARSA to improve charging efficiency. The proposed RLMCA combines dynamic threshold-based charging requests and an improved reward function that accounts for sensor coverage contribution. Extensive simulations demonstrate that RLMCA outperforms conventional methods in terms of charging latency, energy usage efficiency, and network coverage.

Enhancing Retrieval-Augmented Generation with Knowledge Graph-Based Soft-Labeling and Triplet Similarity SBERT

title: Enhancing Retrieval-Augmented Generation with Knowledge Graph-Based Soft-Labeling and Triplet Similarity SBERT abstract: In recent years, generative AI has made significant progress in natural language generation, including applications in customer service. Retrieval-Augmented Generation (RAG) technology has been widely used to enhance the accuracy and relevance of AI-generated responses by integrating external knowledge retrieval. This paper proposes an improved RAG system that employs knowledge graphs and a triplet similarity SBERT framework to refine text retrieval performance. The proposed model introduces soft-label generation for training, optimizing textual representation learning and retrieval quality. Experimental results demonstrate that our method outperforms existing retrieval models in accuracy and efficiency.

Question Answering System Based on Graph Neural Networks and Contrastive Learning Combined with Large Language Models

title: Question Answering System Based on Graph Neural Networks and Contrastive Learning Combined with Large Language Models abstract: In the era of information explosion, question-answering (QA) systems are crucial for efficient information retrieval. However, existing QA models face challenges in knowledge updating, semantic understanding, and computational efficiency. This study proposes a QA system integrating Graph Neural Networks (GNNs), Contrastive Learning, and Large Language Models (LLMs) within a Retrieval-Augmented Generation (RAG) framework. Our method enhances vector representation learning through GNNs and contrastive loss while leveraging RAG for efficient knowledge retrieval. Experimental results demonstrate significant improvements in accuracy and computational efficiency compared to baseline models.

Visual secret sharing for interlaced quick response codes

title: Visual secret sharing for interlaced quick response codes abstract: The quick response codes (QR codes) are commonly seen in our daily lives. They are mainly composed of black or white square blocks, containing the URL or other form of information. We may compose several QR codes together by using interlacing for layered QR codes. With the concept of visual secret sharing (VSS), we may separate the interlaced QR code into several shares, and once received shares is more than the threshold, decoding for layered QR codes can be performed. Even the separated QR codes are not 100% recoverable, they can still be scanned and the information therein can be extracted. Our experiments have presented the applicability of the composition of two QR codes with the concept of visual secret sharing in the proposed method.

An Efficient Cluster-Based Continual Learning with Gradient Episodic Cache Memory

title: An Efficient Cluster-Based Continual Learning with Gradient Episodic Cache Memory

A Novel Time-Awareness Recommendation System

title: A Novel Time-Awareness Recommendation System

應用生成對抗網路於動態社群網路預測

title: 應用生成對抗網路於動態社群網路預測

An Augmented User Context-Awareness Recommendation System

title: An Augmented User Context-Awareness Recommendation System

A Novel RNN-Based Ensemble Model for Link Prediction on Dynamic Social Network

title: A Novel RNN-Based Ensemble Model for Link Prediction on Dynamic Social Network

Error control for content-adaptive block compressive sensing with polar codes and unequal protection concepts

title: Error control for content-adaptive block compressive sensing with polar codes and unequal protection concepts abstract: Block compressive sensing (BCS) is one of the new compression techniques in recent years. In addition to looking for compression performances, protection of BCS plays another key role for robust data transmission. In our scenario, original image can be classified into smooth and active blocks, and then different measurement rates in BCS can be applied. BCS compressed data would be presented with the binary format to form bitstreams for transmitting over binary channels. Considering the concept of unequal error protection, different level of protection can be accomplished with the polar codes. Compressed bistreams protected with polar codes can be transmitted over the binary symmetric channel (BSC), and errors may be induced during transmission. After reception, polar decoding and BCS reconstruction can be utilized subsequently to obtain reconstructed image. Simulations have shown the advantages of using the inherent characteristics of original image and the concept of unequal error protection over conventional settings.

Reversible data hiding based on blockwise prediction with quadtree concepts

title: Reversible data hiding based on blockwise prediction with quadtree concepts abstract: Reversible data hiding is famous for the reversibility in information security researches. For assessing the effectiveness of reversible data hiding schemes, the marked image quality and the payload should be compared, and the reversibility should be guaranteed. It would be advantageous to take the inherent characteristics of original image into consideration to enhance the image quality and embedding payload. By generating predicted image and taking the differences between predicted and original image, a much enhanced amount of payload can be achieved. We employ quadtree decomposition for presenting image characteristics, choose different sets of prediction parameters, and perform embedding with reversible data hiding. Experimental results have provided the flexibility for performance enhancements.

智慧停車場車輛追蹤系統設計

title: 智慧停車場車輛追蹤系統設計

From Handwriting to Calligraphy: A GAN-Based Intuitive System with Cross-Attention and Skeleton Modeling

title: From Handwriting to Calligraphy: A GAN-Based Intuitive System with Cross-Attention and Skeleton Modeling

Can Semantics-Driven Segmentation Improve BERT for Long Legal Texts?

title: Can Semantics-Driven Segmentation Improve BERT for Long Legal Texts?

Taming the Mirage: A Taxonomy of Detection Methods for LLM Hallucinations

title: Taming the Mirage: A Taxonomy of Detection Methods for LLM Hallucinations

Multi-UAV Data Collection with In-Flight Wireless Power Transfer

title: Multi-UAV Data Collection with In-Flight Wireless Power Transfer

Deep-Learning-Based Risk Prediction with Urban Sensing Data for Consumer Driving Safety

title: Deep-Learning-Based Risk Prediction with Urban Sensing Data for Consumer Driving Safety abstract: With the provision of IoT-driven and user-provided sensing data sources in smart cities, we take advantage of deep learning techniques to analyze the spatio-temporal traffic data and predict traffic risks for driving safety. Our study continues to collect the traffic data from multiple sensing sources, and meanwhile adopts both CNN and LSTM to interpret the data collection in spatial and temporal dimensions. Thus, a novel traffic risk prediction scheme based on CNN and LSTM, named TRP-CL, is proposed to generate a traffic warning map of risks and hazard situations on a grid-scaled city map. Not only a theoretic formation but also an experimental implementation of the TRP-CL scheme are made to show the practical feasibility.

SFFTT: A Shared-Parameter and Fast Fourier Transform Lite Transformer

title: SFFTT: A Shared-Parameter and Fast Fourier Transform Lite Transformer abstract: In recent years, Transformer models have achieved remarkable success in natural language processing, yet their enormous number of parameters and high computational complexity restrict their application in resource-constrained environments. This thesis proposes a lightweight Transformer variant, termed SFFTT, which replaces the traditional self-attention mechanism with Fast Fourier Transform (FFT) in the first half of the encoder and employs parameter sharing along with attention threshold filtering in the latter encoder layers and the decoder. Additionally, we introduce SFFTTwithDyT by substituting all Layer Normalization layers with Dynamic Tanh normalization to enhance training stability and model expressiveness. Experimental results demonstrate that the SFFTT series models maintain competitive performance while significantly reducing parameter count and computational cost, offering an effective solution for lightweight Transformer applications.

Multimodal video forgery detection using vision-language and audio features

title: Multimodal video forgery detection using vision-language and audio features

AI-generated image detection based on ViT using frequency domain features

title: AI-generated image detection based on ViT using frequency domain features

DoggyTongue: Automate Tongue Segmentation for TCM Diagnosis

title: DoggyTongue: Automate Tongue Segmentation for TCM Diagnosis

Reversible Network with Meta ActNorm for Content-Preserved Image Style Transfer

title: Reversible Network with Meta ActNorm for Content-Preserved Image Style Transfer abstract: This study focuses on image style transfer techniques and presents improvements to the ArtFlow framework proposed by Jie An et al.. ArtFlow employs a reversible mechanism that maps an image from the pixel space to the feature space during the forward process, transforming it into a feature vector. A style transfer module then converts this feature vector into a stylized one. In the reverse process, the stylized feature vector is mapped back to the pixel space to obtain the final stylized image, effectively preserving the original content details. This reversible design helps prevent content leakage, a common issue in traditional methods. However, there remains room for improvement in terms of adaptability. To enhance the model's adaptability while maintaining structural fidelity, this paper modifies the core activation normalization mechanism within the ArtFlow framework and proposes a novel normalization approach. Inspired by the concept of "learning to learn" in meta-learning, we introduce Meta Activation Normalization (Meta-Actnorm). The improved architecture is termed the Multi-Block Adaptive Flow (MBAF) model. In the MBAF model, Meta-Actnorm dynamically adjusts normalization parameters based on the input image during the forward process and effectively integrates these parameters during the reverse process, further enhancing the model’s adaptability and stability. A series of experiments and quantitative evaluations demonstrate that the proposed method not only preserves key structures and details of the content image but also ensures visual consistency after style transfer, avoiding distortions such as structural deformation.

Research on Performance Improvement of Vision Transformer Model Based on BEiT

title: Research on Performance Improvement of Vision Transformer Model Based on BEiT abstract: Vision Transformer (ViT) has demonstrated exceptional performance in image classification tasks across large-scale datasets. However, its application in domain-specific or small-scale datasets remains a challenge. This research explores an alternative approach to image patch generation, replacing the fixed-size patch mechanism in ViT with semantic-aware segmentation using the Segment Anything Model (SAM). We focus on applying this technique to datasets such as marine biology, animals, and plants, where semantic consistency plays a more critical role. The segmented patches are compared to the conventional 16×16 patches used in ViT to evaluate their potential to enhance semantic representation. Preliminary results suggest that using SAM-based patches can introduce better localized and meaningful features, providing a foundation for performance enhancement in downstream tasks.

Optimizations of Lung Cancer Detection and Classification based on YOLO Architecture

title: Optimizations of Lung Cancer Detection and Classification based on YOLO Architecture abstract: Lung cancer remains a leading cause of cancer-related deaths worldwide. Early detection using chest CT scans can significantly improve patient outcomes, yet accurate diagnosis remains a challenge due to the complex morphology of lung nodules. This paper presents a modified YOLO-based deep learning framework that enhances real-time detection and classification of lung nodules. We introduce architectural changes such as the use of RepC3 modules and deeper convolutional layers in the backbone to improve feature extraction and localization. Experimental results on the LIDC-IDRI dataset show a mean Average Precision (mAP@0.5) of 77.74%, demonstrating the model's effectiveness in detecting and classifying nodules with reasonable accuracy and speed.

Application of Lidar Detection Module in Bicycle Haz-ard Detection and Treatment

title: Application of Lidar Detection Module in Bicycle Haz-ard Detection and Treatment

從判決到預測：台灣法院判決文之車禍慰撫金建模前處理研究

title: 從判決到預測：台灣法院判決文之車禍慰撫金建模前處理研究

PGNet v2: Enhancing Aesthetic Image Critique Generation with Self-Resurrecting Activation and Gaussian Gated Units

title: PGNet v2: Enhancing Aesthetic Image Critique Generation with Self-Resurrecting Activation and Gaussian Gated Units

Vehicle Tracking Using Lane Cameras for Personalized Parking Guidance

title: Vehicle Tracking Using Lane Cameras for Personalized Parking Guidance

Image Outpainting Based On Attention Model

title: Image Outpainting Based On Attention Model abstract: Along the advanced progresses on deep neural networks, there are many impressive results on image inpainting. Consequently, several research are trying to transfer successful experiences into image outpainting. Contextual attention net is one of the popular architectural units being applied to outpainting. We argue that it may not as suitable when embedded in an outpainting network. Instead, we adopt SEnet for it has global receptive field and channel-wise feature recalibration. This is very helpful for image outpainting. We also propose a local discriminator mechanism to decide whether a randomly select partial image is a real one. By ‘randomness’, the generator can produce a realistic result.

Deep-Learning-Based Risk Prediction with Urban Sensing Data for Consumer Driving Safety

title: Deep-Learning-Based Risk Prediction with Urban Sensing Data for Consumer Driving Safety