댓글 0건 조회 57회 작성일 2023-09-16 14:05
- Professor Sang Hyun Park's team developed an image conversion model that removes data bias.
- Learning a deep learning model with incomplete data can perform well on data from a variety of sources, enabling the development of a strong deep learning model in healthcare and industry where data collection is difficult on the Internet and in various institutions.
- The proposed image conversion model first extracts texture information from the content information of the input image and the image of the new domain, and inputs these two into the generation model to generate the image. To maintain the content information of the input image and the texture information of the new domain, the texture co-occurrence error function and the content self-similarity error function are used for learning.
- Using the proposed technique, we confirm that the performance improves significantly when the deep learning model is learned after building a dataset with removed texture bias. In addition, through ablation research, it was confirmed that the proposed technique is qualitatively and quantitatively superior to the technique of maintaining other content or transforming the texture. This study is published in Neural Networks (IF=7.8), the top international academic journal in the field of artificial intelligence.