Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Jaepil Ko has a total of 19 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Anti-Aliasing Framework for Time-Domain Speech Enhancement

    by Joonhyeon Bae, Jaepil Ko

    Since the introduction of deep learning, research on speech enhancement has been developing rapidly. Until a few years ago, time-domain models performed poorly compared to time-frequency models. However, recent time-domain models such as DEMUCS, MANNER, and SE-Conformer have achieved competitive performance by adopting U-Net encoder-decoder architectures with skip connections. Despite this progress, these U-Net-based models have an overlooked weakness: aliasing artifacts from down-sampling and up-sampling operations. Prior work has shown that max pooling and strided convolutions introduce aliasing, which appears as network-generated high-frequency noise in the output. This contradicts the fundamental goal of speech enhancement noise reduction. Therefore, applying anti-aliasing methods can be useful for further improving time-domain models. We systematically apply anti-aliasing methods (BlurPooling, sub-pixel convolution with ICNR initialization) to MANNER, a time-domain speech enhancement model, on the VoiceBank+DEMAND dataset. Our results show improvements in perceptual quality metrics. Our contributions are threefold. First, we investigate aliasing artifacts caused by down-sampling and up-sampling operations in U-Net-based time-domain speech enhancement models. Second, we formulate a unified 1D anti-aliasing framework that integrates existing anti-aliasing operations, including BlurPooling, sub-pixel convolution, and ICNR initialization, into waveform-domain speech enhancement architectures. Third, we provide empirical validation on the framework through objective metrics, ablation studies, an additional Demucs experiment, and spectrogram/error-spectrogram visualizations.

  • Memory-Induced Attractor Dynamics in Liquid Time-Constant Networks

    by Bishal Ranjan Swain, Jaeyeon Hwang, Jaepil Ko

  • Open Access

    Unified Adult–Pediatric Glioma Segmentation via Synergistic MAE Pretraining and Boundary-Aware Refinement

    by Moldir Zharylkassynova, Jaepil Ko, Kyungjoo Cheoi

    Accurate brain tumor segmentation in both adult and pediatric populations remains a challenge due to substantial differences in brain anatomy, tumor distribution, and subregion size. This study proposes a unified segmentation framework based on nnU-Net, integrating encoder-level self-supervised pretraining with a lightweight, boundary-aware decoder. The encoder is initialized using a large-scale 3D masked autoencoder pretrained on brain MRI, while the decoder is trained with a hybrid loss function that combines region-overlap and boundary-sensitive terms. A harmonized training and evaluation protocol is applied to both the BraTS-GLI (adult) and BraTS-PED (pediatric) cohorts, enabling fair cross-cohort comparison against baseline and advanced nnU-Net variants. The proposed method improves mean Dice scores from 0.76 to 0.90 for adults and from 0.64 to 0.78 for pediatric cases, while reducing HD95 from 4.42 to 2.24 mm and from 9.03 to 6.23 mm, respectively. These results demonstrate that combining encoder-level pretraining with decoder-side boundary supervision significantly enhances segmentation accuracy across age groups without adding inference-time computational overhead.


  • Open Access

    Deep Learning-Based Allergic Rhinitis Diagnosis Using Nasal Endoscopy Images

    by Jaepil Ko, MinHye Kang, Young Joon Jun

    Purpose Allergic rhinitis typically has edematous and pale turbinates or erythematous and inflamed turbinates. While traditional approaches include using skin prick tests (SPT) to determine the presence of AR, It is often not related to actual symptoms, and it is an invasive test. We use deep learning to analyze nasal endoscopy images to investigate a quantitative method for diagnosing allergic rhinitis. Methods Traditional machine learning-based diagnostic techniques have relied on structured clinical datasets featuring statistical data such as demographic characteristics, symptom severity, and clinical test results. In contrast, we propose a novel approach to use endoscopy image data to analyze the color distribution in the inferior turbinate region of patients with allergic rhinitis using the CIE-Lab color space and extract the adaptive histogram features that are used to explore and find suitable feature extraction methods and deep learning model architectures. Results Our proposed model achieves a promising diagnostic accuracy of 90.80% for images exhibiting AR symptoms. Future research will expand the dataset to include a broader spectrum of symptomatic and asymptomatic images to enhance model robustness and investigate the potential of optical analysis as a non-invasive diagnostic method for AR. Conclusion This study introduced a novel approach to diagnosing allergic rhinitis using nasal endoscopy images. Our approach analyzed the color distribution of the inferior turbinates within the LAB color space, extracted important features from endoscopy images using both CNN feature extraction and histograms, and performed classification through SVM and fully connected classifiers.

  • Open Access

    Enhancing Nuclei Segmentation in Histology Images by Effectively Merging Task-Specific and Foundation Models

    by Bishal Ranjan Swain, Kyung Joo Cheoi, Jaepil Ko

    한국멀티미디어학회(The Journal of Curriculum Evaluation)

  • Open Access

    Leveraging Multi-Modality and Enhanced Temporal Networks for Robust Violence Detection

    by Gwangho Na, Jaepil Ko, Kyungjoo Cheoi

    In this paper, we present a novel model that enhances performance by extending the dual-modality TEVAD model—originally leveraging visual and textual information—into a multi-modal framework that integrates visual, audio, and textual data. Additionally, we refine the multi-scale temporal network (MTN) to improve feature extraction across multiple temporal scales between video snippets. Using the XD-Violence dataset, which includes audio data for violence detection, we conduct experiments to evaluate various feature fusion methods. The proposed model achieves an average precision (AP) of 83.9%, surpassing the performance of single-modality approaches (visual: 73.9%, audio: 67.1%, textual: 29.9%) and dual-modality approaches (visual + audio: 78.8%, visual + textual: 78.5%). These findings demonstrate that the proposed model outperforms models based on the original MTN and reaffirm the efficacy of multi-modal approaches in enhancing violence detection compared to single- or dual-modality methods.


  • Open Access

    Enhancing Surveillance Systems: Integration of Object, Behavior, and Space Information in Captions for Advanced Risk Assessment

    by Minseong Jeon, Jaepil Ko, Kyungjoo Cheoi

    This paper presents a novel approach to risk assessment by incorporating image captioning as a fundamental component to enhance the effectiveness of surveillance systems. The proposed surveillance system utilizes image captioning to generate descriptive captions that portray the relationship between objects, actions, and space elements within the observed scene. Subsequently, it evaluates the risk level based on the content of these captions. After defining the risk levels to be detected in the surveillance system, we constructed a dataset consisting of [Image-Caption-Danger Score]. Our dataset offers caption data presented in a unique sentence format, departing from conventional caption styles. This unique format enables a comprehensive interpretation of surveillance scenes by considering various elements, such as objects, actions, and spatial context. We fine-tuned the BLIP-2 model using our dataset to generate captions, and captions were then interpreted with BERT to evaluate the risk level of each scene, categorizing them into stages ranging from 1 to 7. Multiple experiments provided empirical support for the effectiveness of the proposed system, demonstrating significant accuracy rates of 92.3%, 89.8%, and 94.3% for three distinct risk levels: safety, hazard, and danger, respectively.


  • Open Access

    Complex-Phase Steel Microstructure Segmentation Using UNet: Analysis across Different Magnifications and Steel Types

    by Bishal Ranjan Swain, Dahee Cho, Joongcheul Park, Jae-Seung Roh, Jaepil Ko

    The quantification of the phase fraction is critical in materials science, bridging the gap between material composition, processing techniques, microstructure, and resultant properties. Traditional methods involving manual annotation are precise but labor-intensive and prone to human inaccuracies. We propose an automated segmentation technique for high-tensile strength alloy steel, where the complexity of microstructures presents considerable challenges. Our method leverages the UNet architecture, originally developed for biomedical image segmentation, and optimizes its performance via careful hyper-parameter selection and data augmentation. We employ Electron Backscatter Diffraction (EBSD) imagery for complex-phase segmentation and utilize a combined loss function to capture both textural and structural characteristics of the microstructures. Additionally, this work is the first to examine the scalability of the model across varying magnifications and types of steel and achieves high accuracy in terms of dice scores demonstrating the adaptability and robustness of the model.