다중 스케일 이산 웨이블릿 변환을 통한 잠재 확산 모델 기반 초해상화의 세부 품질 향상

초록

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 models. Existing SR approaches that utilize latent diffusion models have struggled to preserve fine details, such as subtle contours and textures, during the super-resolution process of low-resolution images. In this study, both the super-resolved images generated by the model and the corresponding high-resolution images are augmented across various scales. The errors between them are then extracted in the high-frequency domain using discrete wavelet transform. This approach enables the model to retain detailed information more accurately. Experimental results on benchmark datasets demonstrate that the proposed framework outperforms previous methods, effectively improving the detail quality of super-resolved images.

키워드