Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hong-Yi Chang has a total of 46 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Cross-Domain Robust Pruning for Polyp Segmentation: Multi-Encoder Feature Fusion Beats Single-Encoder Baselines

    by Chia-Pei Tang, Hong-Yi Chang, Tzu-Shan Chang, Yu-Chieh Chang, Chia-Hsin Cheng

    Medical image segmentation requires dense pixel-level annotations, making large-scale dataset construction expensive and motivating research into data-efficient training. The main objective of this paper is to determine whether fusing two complementary pretrained image encoders into a single similarity space can make training-free dataset pruning robust across heterogeneous polyp segmentation domains and to quantify that robustness against a comprehensive panel of baselines. To achieve this, we propose Multi-Encoder Diverse Pruning (MEDP), a training-free dataset-pruning method. MEDP fuses features from an ImageNet-pretrained ResNet-18 and a self-supervised DINOv2 ViT-S/14 into a single 896-D similarity space. It partitions the training pool via Louvain modularity maximization and selects per-community samples via maximal-marginal-relevance (MMR) ranking, which effectively balances eigenvector centrality with feature-space diversity. We benchmarked MEDP against 12 baselines at a 20% retention ratio across three polyp segmentation settings (Kvasir-SEG, CVC-ClinicDB, and a Combined cross-domain pool) using a standard 5-level UNet. Based on approximately 330 controlled training runs, the results demonstrate that MEDP achieves the highest mean test Dice of 0.7324 on the most challenging Combined cross-domain pool, significantly outperforming uniform random sampling (Cohen’s d = +1.79, paired Wilcoxon p = 0.002). Conversely, all hand-crafted structure-aware variants failed to outperform uniform random sampling. These findings confirm that combining multi-encoder features with MMR diversity provides a simple and effective strategy for improving robustness across heterogeneous medical imaging settings and that the choice of pretrained image encoder is the dominant factor in segmentation-aware pruning.

  • Open Access

    Ordinal Regression Research Based on Dual Loss Function—An Example on Lumbar Vertebra Classification in CT Images

    by Chia-Pei Tang, Hong-Yi Chang, Yu-Ming Hsu, Tu-Liang Lin

    Background/Objectives: Some classification problems involve ordered categories (e.g., low–medium–high), which are better modeled as ordinal regression. This study aimed to propose and evaluate a dual loss framework—Ordinal Residual Dual Loss—for lumbar vertebra classification on CT images to assist L3 identification and sarcopenia detection. Methods: In this retrospective study, lumbar spine CT images were used to train a deep learning model based on a MobileNet-v3-Large network. The proposed framework combines standard cross-entropy loss for classification with an Ordinal Residual Loss defined on the difference between output probabilities and target ordinal probabilities. Results: Experimental results show that the Ordinal Residual Dual Loss approach outperforms using cross-entropy alone and also surpasses methods from previous studies in lumbar vertebra classification performance. Conclusions: Leveraging a dual loss design that incorporates ordinal information improves vertebral level classification on CT images and has potential to support more accurate automated L3 localization and sarcopenia assessment in clinical practice.

  • Open Access

    Massive open online course recommendation system based on a reinforcement learning algorithm

    by Jian-Wei Tzeng, Nen-Fu Huang, An-Chi Chuang, Ting-Wei Huang, Hong-Yi Chang

    Massive open online courses (MOOCs) are open online courses designed on the basis of the teaching progress. Videos and learning exercises are used as learning materials in these courses, which are open to numerous users. However, determining the prerequisite knowledge and learning progress of learners is difficult. On the basis of learners’ online learning trajectory, we designed a set of practice questions for a recommendation system for MOOCs, provided suitable practice questions to students through the LINE chatbot (a type of social media software), and used mobile devices to encourage participation in MOOCs. Reinforcement learning, which involves reward function design and iterative solution improvement, was used to set task goals, including those related to course learning and practice question difficulty. The proposed system encouraged certain learning behaviors among students. Students who used the system exhibited an exercise completion rate of 89.97%, which was higher than that of students who did not use the system (47.23%). The system also increased the students’ overall learning effectiveness. Students who used and did not use the proposed system exhibited average midterm scores of 64.73 and 58.21, respectively. We also collected 227 online questionnaires from students. The results of the questionnaires indicated that 90% of the students were satisfied with the system and hoped to continue using it.


  • Ransomware Detection by Distinguishing API Call Sequences through LSTM and BERT Models

    by Tu-Liang Lin, Hong-Yi Chang, Yuan-Yao Chiang, Shu-Cheng Lin, Tsung-Yen Yang, Chun-Jun Zhuang, Wha-Lee Tseng, Bo-Hao Zhang

    Nowadays, ransomware evolved rapidly and the prevention of ransomware has become an important issue. The threat of ransomware is much more sophisticated than before for governments and enterprises; breaches or corruption of sensitive data will cause huge impact on the organization. Early detection is one of the effective method to prevent the ransomware attack. Modern ransomware detection technologies can be divided into two categories: static analysis and dynamic analysis. Dynamic analysis observes the behavior of the running program. Previous research adopted machine learning approach for dynamic analysis and API sequence dataset were used to trained machine learning models for dynamic analysis. In this research, we collected the API calls of the ransomware from reports generated by Cuckoo Sandbox and proposed two detecting models using BERT and LSTM. The result shows that both BERT and LSTM models can successfully predict ransomware with 95% high accuracy. We aimed to compare the performance of two text-based learning model, LSTM and BERT, and analyze the pros and cons. The result shows that API sequence data can be used to train effective ransomware detection models in text-based manner.


  • Open Access

    A Novel Computer-Aided Detection/Diagnosis System for Detection and Classification of Polyps in Colonoscopy

    by Chia-Pei Tang, Hong-Yi Chang, Wei-Chun Wang, Wei-Xuan Hu

    Using a deep learning algorithm in the development of a computer-aided system for colon polyp detection is effective in reducing the miss rate. This study aimed to develop a system for colon polyp detection and classification. We used a data augmentation technique and conditional GAN to generate polyp images for YOLO training to improve the polyp detection ability. After testing the model five times, a model with 300 GANs (GAN 300) achieved the highest average precision (AP) of 54.60% for SSA and 75.41% for TA. These results were better than those of the data augmentation method, which showed AP of 53.56% for SSA and 72.55% for TA. The AP, mAP, and IoU for the 300 GAN model for the HP were 80.97%, 70.07%, and 57.24%, and the data increased in comparison with the data augmentation technique by 76.98%, 67.70%, and 55.26%, respectively. We also used Gaussian blurring to simulate the blurred images during colonoscopy and then applied DeblurGAN-v2 to deblur the images. Further, we trained the dataset using YOLO to classify polyps. After using DeblurGAN-v2, the mAP increased from 25.64% to 30.74%. This method effectively improved the accuracy of polyp detection and classification.


  • Open Access

    A Novel Unmanned Aerial Vehicle Charging Scheme for Wireless Rechargeable Sensor Networks in an Urban Bus System

    by Tu-Liang Lin, Hong-Yi Chang, Yu-Hsin Wang

    Wireless sensor networks (WSNs) are implemented in many aspects of daily life, such as Internet of Things applications, industrial automation, and intelligent agriculture. Sensors are typically powered by batteries. Chargers can be used to supply power to sensor nodes and thus extend the lifetime of WSNs. This special type of network is named a wireless rechargeable sensor network (WRSN). However, due to the limited battery power and different deployment locations of the sensors, efficiently moving the chargers from the current sensor nodes to the next sensor nodes is a challenge. In this study, we propose an unmanned aerial vehicle (UAV)-based charging scheme in an urban bus system, involving the coordination between UAVs and bus schedules. The UAVs can be recharged by urban buses and then supply the power to sensor nodes. We implemented three charging strategies: naïve, shortest path, and max power. In the naïve strategy, the UAVs fly directly to sensor nodes when the sensors are lacking power. In the shortest path strategy, the minimum distance between the sensor node and bus location is calculated, and the UAVs fly the shortest path to the sensor nodes. In the maximum power charging strategy, the UAV that has the highest battery power is assigned to work. The experimental results show that the shortest path charging and max power charging strategies perform better than naïve charging in different parameter settings. To prolong the lifetime of the network system, adjusting the bus frequency according to the number of nearby sensors around the bus route is favorable.


  • Open Access

    A Novel Hybrid Search and Remove Strategy for Power Balance Wireless Charger Deployment in Wireless Rechargeable Sensor Networks

    by Tu-Liang Lin, Hong-Yi Chang, Yu-Hsin Wang

    Conventional sensor nodes are often battery-powered, and battery power limits the overall lifetime of the wireless sensor networks (WSNs). Wireless charging technology can be implemented in WSNs to supply power to sensor nodes and resolve the problem of restricted battery power. This type of mixed network is called wireless rechargeable sensor networks (WRSNs). Therefore, wireless charger deployment is a crucial task in WRSNs. In this study, the method of placing wireless chargers to efficiently extend the lifetime of the WRSNs is addressed. Owing to the data forwarding effect in WSNs, sensor nodes that are closer to the data collection or sink node drain more power than nodes that are further away from the data collection or sink node. Therefore, this study proposes a novel hybrid search and removal strategy for the power balance charger deployment method. The wireless chargers are placed in the chosen nodes of the WRSNs. The node-chosen problem we address is called the dominating set problem. The proposed hybrid search and removal strategy attempts to discover the minimum number of chargers required to cover all sensor nodes in the WRSN. The proposed algorithm considers the charging power of the wireless directional charger when arranging its placement to maximize the charging capacity in a power-balanced prerequisite. Therefore, the proposed deployment strategy preserves the awareness of the presence of the sink node that could result in unbalanced power distribution in WRSNs. The simulation results show that the proposed strategy spares more chargers and achieves better energy efficiency than other deployment approaches.


  • Open Access

    A Location Privacy Attack Based on the Location Sharing Mechanism with Erroneous Distance in Geosocial Networks

    by Tu-Liang Lin, Hong-Yi Chang, Sheng-Lin Li

    Geographical social networks (GSN) is an emerging research area. For example, Foursquare, Yelp, and WeChat are all well-known service providers in this field. These applications are also known as location-based services (LBS). Previous studies have suggested that these location-based services may expose user location information. In order to ensure the privacy of the user’s location data, the service provider may provide corresponding protection mechanisms for its applications, including spatial cloaking, fuzzy location information, etc., so that the user’s real location cannot be easily cracked. It has been shown that if the positioning data provided by the user is not accurate enough, it is still difficult for an attacker to obtain the user’s true location. Taking this factor into consideration, our attack method is divided into two stages for the entire attack process: (1) Search stage: cover the area where the targeted user is located with unit discs, and then calculate the minimum dominating set. Use the triangle positioning method to find the minimum precision disc. (2) Inference phase: Considering the existence of errors, an Error-Adjusted Space Partition Attack Algorithm (EASPAA) was proposed during the inference phase. Improved the need for accurate distance information to be able to derive the user’s true location. In this study, we focus on the Location Sharing Mechanism with Maximal Coverage Limit to implement the whole attack. Experimental results show that the proposed method still can accurately infer the user’s real location even when there is an error in the user’s location information.


  • IPARBC: An Improved Parallel Association Rule Based on MapReduce Framework

    by Hong-Yi Chang, Zih-Huan Hong, Tu-Liang Lin, Wan-Kun Chang, Yi-Ying Lin

    Recently, data mining in big data has become an important concern for researchers. Data mining, which refers to mining the relationship between items in a dataset, has been applied by businesses to seek profitable outcomes. Association rule mining algorithms such as Apriori and the FP-growth are efficient methods for discovering relations between items in large databases. To enhance the performance, many researches tend to enhance the traditional method using the MapReduce framework. In this paper, we proposed an improved association rule algorithm (IPARBC) based on MapReduce framework. The concept of combinatorial mathematics is used as the theoretical basis of the algorithm, and in order to improve mining performance by MapReduce framework, we address the high volume problem of big data. Experimental results show that the proposed algorithm outperform other algorithms substantially in terms of runtime.