Jun-Hyeok Seo

In Preparation

Under Review

This paper proposes a context-aware adaptive mask selection framework for referring video object segmentation (RVOS). The method integrates CLIP-based keyframe selection, mask consistency scoring, text-guided object number determination, and temporal mask interpolation to ensure both temporal stability and accurate multi-object segmentation.

This paper proposes a multi-scale discrete wavelet transform framework designed to enhance the detail quality of super-resolution (SR) methods based on latent diffusion model. Experimental results on benchmark datasets demonstrate that the proposed framework outperforms previous methods, effectively improving the detail quality of super-resolved images.

This paper proposes a multi-perspective prompt-based framework for zero-shot image classification using LVLMs, enhancing robustness by integrating object-aware, intent-aware, and reason-aware prompts, and achieving superior performance through ensemble prediction on the VizWiz dataset.

This paper proposes a diffusion model-based image super-resolution framework using multi-scale high-frequency error maps, enabling accurate restoration of image details and outperforming existing methods on benchmark datasets.

This paper proposes a robust OCR algorithm using the end-to-end Transformer-based TrOCR model, achieving a Character Error Rate (CER) of approximately 0.08 and significantly outperforming Tesseract OCR and EasyOCR.

This paper proposes a system for detecting vehicle damage areas through semantic segmentation using the UNet++ model, achieving pixel-level accuracy with an average IoU of 0.94 and comparing its performance with the conventional U-Net model.

This paper proposes a deep learning-based technique for detecting stains in clothing images using a UNet++ model with an EfficientNet encoder pre-trained on ImageNet, achieving an average IoU of approximately 0.9589.

This paper analyzes various words used in CQI reports and constructs a domain dictionary to provide the foundation for future automatic analysis of such reports.

This paper proposes a vehicle model recognition and license plate identification system using YOLOv5, enabling both tasks simultaneously and reducing computational cost compared to previous methods.