This project developed a multilingual toxicity detection system for Finnish and German using data augmentation via machine translation and multilingual transfer learning with XLM-RoBERTa.

This project developed a multilingual toxicity detection system for Finnish and German using data augmentation via machine translation and multilingual transfer learning with XLM-RoBERTa, significantly improving performance on low-resource languages. The approach integrated diverse datasets and majority voting, achieving a top ranking in a multilingual toxicity detection competition.

This project visualizes human pose estimation by synchronizing original video with both 2D and 3D skeletal joint tracking and enables intuitive analysis of complex movement patterns.

This project visualizes human pose estimation by synchronizing original video with both 2D and 3D skeletal joint tracking and enables intuitive analysis of complex movement patterns. By integrating MediaPipe's advanced joint detection, it offers a comprehensive side-by-side comparison of motion from multiple perspectives.

This project investigates automatic speech recognition for Gibberish EsperantO using wav2vec2 models implemented via the Hugging Face Transformers library.

This project investigates automatic speech recognition for Gibberish EsperantO (an Esperanto variant with systematic character substitutions) using wav2vec2 models implemented via the Hugging Face Transformers library. By fine-tuning multilingual and Esperanto-specific wav2vec2 architectures, the study demonstrates the effectiveness and limitations of self-supervised deep learning for robust ASR in heavily distorted transcription scenarios.

This project explores deep learning-based methods for precise segmentation of coronary vessels in X-ray angiogram images using UNet, CapsuleNet, and GraphUNet models.

This project explores deep learning-based methods for precise segmentation of coronary vessels in X-ray angiogram images. By comparing models such as UNet, CapsuleNet, and GraphUNet with advanced preprocessing techniques like CLAHE, the research aims to enhance segmentation accuracy and robustness. The resulting algorithms enable automatic quantification and categorization of coronary artery stenosis to support clinical diagnosis.