Chun-An Lin

Prof. Chun-An Lin

orcid

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
19 May 2024
Verified
Conference Paper
Closed access
CLChun-An Lin
Chun-An Lin
TLTsung-Jung Liu
Tsung-Jung Liu
KLKuan-Hsien Liu
Kuan-Hsien Liu
Infrared imaging is crucial for many applications, such as night vision and environmental monitoring. However, it often suffers from lower spatial resolution compared to visible light imaging, which hinders its effectiveness in scenes demanding fine detail discernment. This paper introduces a Lightweight Infrared Image Super-Resolution Network (LIRSRN), which is a novel architecture designed to enhance the resolution of IR images with minimal computation overhead. The core of LIRSRN is the Attention Enhancement Module, which synergizes various attention mechanisms to emphasize important features across both channel and spatial domains. It is further enhanced by spatial frequency processing, which helps to extract key image attributes. The training is conducted on the DIV2K dataset, and the test results from "results-A" and "results-C" demonstrate the superior performance of LIRSRN in terms of PSNR and SSIM. Ablation studies compare the trade-offs between performance and computational cost using different attention mechanisms. Overall, the proposed model strikes a balance between performance and complexity. The source code and trained model are available at https://reurl.cc/yYDnA6.

16 September 2025 in IEEE Access
Verified
Journal Article
Open access
YYYuan-Cheng Yu
Yuan-Cheng Yu
YOYen-Chieh Ouyang
Yen-Chieh Ouyang
CLChun-An Lin
Chun-An Lin
The escalating diversity of Internet of Things (IoT) devices has intensified security challenges, as signature-based and heavyweight machine learning (ML) defenses often fall short in resource-constrained edge environments, unable to meet real-time demands. To overcome these limitations, we introduce Patch Gate Time-series Anomaly Detection (PGTAD), a novel, lightweight, unsupervised multivariate time-series (MVTS) anomaly detection algorithm specifically tailored for IoT network traffic. PGTAD is a Gated recurrent units (GRU) autoencoder-based framework that intelligently integrates features from three perspectives—forward sequences, backward sequences, and local patch segments—using a Gate Mechanism (GM). It employs sinusoidal positional encoding for temporal awareness and a patch operation to capture localized patterns. Anomaly detection is achieved via Mahalanobis distance on reconstruction errors. Extensive experiments on publicly available IoT network traffic datasets: Edge-IIoTset, CICIoT2023, CICIoMT2024, and CICIoV2024, utilizing both packet-based and flow-based feature perspectives, demonstrated PGTAD's superior performance. Evaluated using various F-score metrics (including point-wise, composite, and affiliation) and inference throughput, PGTAD achieved an average F-score of 0.92 and sustained over 480K packets per second on an NVIDIA Jetson Orin Nano 8GB development kit. This confirms its effectiveness and generalizability for real-time detection in IoT edge environments.

2 July 2024
Verified
Conference Paper
Closed access
YYYuan-Cheng Yu
Yuan-Cheng Yu
YOYen-Chieh Ouyang
Yen-Chieh Ouyang
LWLing-Wei Wu
Ling-Wei Wu
CLChun-An Lin
Chun-An Lin
KTKuo-Yu Tsai
Kuo-Yu Tsai
Effective anomaly detection is vital to minimize the economic impacts of security issues that Internet of Things (IoT) and Industrial Internet of Things (IIoT) face more frequently in recent years. Conventional anomaly detection methods have difficulties in finding new and complex anomalies. machine learning (ML) algorithms have proven to be excellent in anomaly detection with the development of ML research. However, labeled data is hard to get in real situations, so we need unsupervised learning methods. We suggest a new unsupervised learning method for detecting anomalies in multivariate time series, named Bi-Dual-GM GRU-AE. It has two Gate Mechanisms to better pick the combination part from different aspects of input features. These features are derived from different hidden sizes of GRU encoders that take both forward and backward time windows as inputs. We evaluated the proposed method on real-world IoT and IIoT datasets and showed that it outperforms the latest methods in the multivariate time series anomaly detection task. The proposed approach shows superior detection performance particularly in scenario where anomalies are relatively rare, offering a better detection solution to the modern challenges in IoT and IIoT security.

23 September 2025 in IEEE Access
Verified
Journal Article
Open access
YYYuan-Cheng Yu
Yuan-Cheng Yu
YOYen-Chieh Ouyang
Yen-Chieh Ouyang
CLChun-An Lin
Chun-An Lin
Time-series anomaly detection plays a central role across a wide range of application domains. With the increasing proliferation of the Internet of Things (IoT) and smart manufacturing, time-series data has dramatically increased in both scale and dimensionality. This growth has exposed the limitations of traditional statistical methods in handling the high heterogeneity and complexity of such data. Inspired by the recent success of large language models (LLMs) in multimodal tasks across language and vision domains, we propose a novel unsupervised anomaly detection framework: A Tri-Branch Patch-wise Large Language Model Framework for Time-Series Anomaly Detection (TriP-LLM). TriP-LLM integrates local and global temporal features through a triple-branch design comprising Patching, Selecting, and Global modules, to encode the input time-series into patch-wise representations, which are then processed by a frozen, pretrained LLM. A lightweight patch-wise decoder reconstructs the input, from which anomaly scores are derived. We evaluate TriP-LLM on several public benchmark datasets using PATE, a recently proposed threshold-free evaluation metric, and conduct all comparisons within a unified open-source framework to ensure fairness. Experimental results show that TriP-LLM consistently outperforms recent state-of-the-art (SOTA) methods across all datasets, demonstrating strong detection capabilities. Furthermore, through extensive ablation studies, we verify the substantial contribution of the LLM to the overall architecture. Compared to LLM-based approaches using Channel Independence (CI) patch processing, TriP-LLM achieves significantly lower memory consumption, making it more suitable for GPU memory-constrained environments. All code and model checkpoints of TriP-LLM are publicly available on https://github.com/YYZStart/TriP-LLM.git.

12 June 2025 in Diagnostics
Verified
Journal Article
Open access
PDF file
SCSi-Wa Chan
Si-Wa Chan
CLChun-An Lin
Chun-An Lin
YOYen-Chieh Ouyang
Yen-Chieh Ouyang
GCGuan-Yuan Chen
Guan-Yuan Chen
Chein-I ChangChein-I Chang
Chein-I Chang
MLMing-Cheng Liu
Ming-Cheng Liu
Background/Objectives: This research presents a novel analytical method for breast tumor characterization and tissue classification by leveraging intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI) combined with hyperspectral imaging techniques and deep learning. Traditionally, dynamic contrast-enhanced MRI (DCE-MRI) is employed for breast tumor diagnosis, but it involves gadolinium-based contrast agents, which carry potential health risks. IVIM imaging extends conventional diffusion-weighted imaging (DWI) by explicitly separating the signal decay into components representing true molecular diffusion (D) and microcirculation of capillary blood (pseudo-diffusion or D*). This separation allows for a more comprehensive, non-invasive assessment of tissue characteristics without the need for contrast agents, thereby offering a safer alternative for breast cancer diagnosis. The primary purpose of this study was to evaluate different methods for breast tumor characterization using IVIM-DWI data treated as hyperspectral image stacks. Dice similarity coefficients and Jaccard indices were specifically used to evaluate the spatial segmentation accuracy of tumor boundaries, confirmed by experienced physicians on dynamic contrast-enhanced MRI (DCE-MRI), emphasizing detailed tumor characterization rather than binary diagnosis of cancer. Methods: The data source for this study consisted of breast MRI scans obtained from 22 patients diagnosed with mass-type breast cancer, resulting in 22 distinct mass tumor cases analyzed. MR images were acquired using a 3T MRI system (Discovery MR750 3.0 Tesla, GE Healthcare, Chicago, IL, USA) with axial IVIM sequences and a bipolar pulsed gradient spin echo sequence. Multiple b-values ranging from 0 to 2500 s/mm2 were utilized, specifically thirteen original b-values (0, 15, 30, 45, 60, 100, 200, 400, 600, 1000, 1500, 2000, and 2500 s/mm2), with the last four b-value images replicated once for a total of 17 bands used in the analysis. The methodology involved several steps: acquisition of multi-b-value IVIM-DWI images, image pre-processing, including correction for motion and intensity inhomogeneity, treating the multi-b-value data as hyperspectral image stacks, applying hyperspectral techniques like band expansion, and evaluating three tumor detection methods: kernel-based constrained energy minimization (KCEM), iterative KCEM (I-KCEM), and deep neural networks (DNNs). The comparisons were assessed by evaluating the similarity of the detection results from each method to ground truth tumor areas, which were manually drawn on DCE-MRI images and confirmed by experienced physicians. Similarity was quantitatively measured using the Dice similarity coefficient and the Jaccard index. Additionally, the performance of the detectors was evaluated using 3D-ROC analysis and its derived criteria (AUCOD, AUCTD, AUCBS, AUCTD−BS, AUCODP, AUCSNPR). Results: The findings objectively demonstrated that the DNN method achieved superior performance in breast tumor detection compared to KCEM and I-KCEM. Specifically, the DNN yielded a Dice similarity coefficient of 86.56% and a Jaccard index of 76.30%, whereas KCEM achieved 78.49% (Dice) and 64.60% (Jaccard), and I-KCEM achieved 78.55% (Dice) and 61.37% (Jaccard). Evaluation using 3D-ROC analysis also indicated that the DNN was the best detector based on metrics like target detection rate and overall effectiveness. The DNN model further exhibited the capability to identify tumor heterogeneity, differentiating high- and low-cellularity regions. Quantitative parameters, including apparent diffusion coefficient (ADC), pure diffusion coefficient (D), pseudo-diffusion coefficient (D*), and perfusion fraction (PF), were calculated and analyzed, providing insights into the diffusion characteristics of different breast tissues. Analysis of signal intensity decay curves generated from these parameters further illustrated distinct diffusion patterns and confirmed that high cellularity tumor regions showed greater water molecule confinement compared to low cellularity regions. Conclusions: This study highlights the potential of combining IVIM-DWI, hyperspectral imaging techniques, and deep learning as a robust, safe, and effective non-invasive diagnostic tool for breast cancer, offering a valuable alternative to contrast-enhanced methods by providing detailed information about tissue microstructure and heterogeneity without the need for contrast agents.

31 October 2023
Verified
Conference Paper
Closed access
YHYu-Fang Huang
Yu-Fang Huang
TLTsung-Jung Liu
Tsung-Jung Liu
CLChun-An Lin
Chun-An Lin
KLKuan-Hsien Liu
Kuan-Hsien Liu
Nowadays, road object detection has become a challenging topic and a long-standing problem. In previous algorithms, traditional convolutional neural networks (CNNs) have almost dominated computer vision and achieved considerable successes in different detection tasks. However, Transformer-based models in the past two years have shown impressive performance, and some of which even surpass CNN-based methods and become the state-of-the-art for advanced vision tasks. In this paper, we propose a detection model named SOAda-YOLOR, which uses the current mainstream Bi-FPN concept with the innovative Optimized-NMS. We also propose a novel ShuffleSimAM module to the backbone, and apply some bag-of-freebies such as SIoU loss to road object detection, and finally achieve competitive results.