DGIST's Study

A research team led by Professor Sang-hyun Park at Daegu Gyeongbuk Institute of Science and Technology (DGIST) 
has developed an innovative image translation model aimed at mitigating biases in data used for artificial intelligence (AI) models. 
In the development of AI models using images from diverse sources, biases can emerge, affecting model performance. 
The newly created model effectively removes biases without specific information about the factors causing them, enhancing image analysis accuracy. 
This breakthrough is expected to have applications in self-driving technology, content creation, and medical fields.

The problem addressed by the research lies in biases that often arise in datasets used to train deep learning models. 
For example, when creating a dataset to differentiate between bacterial pneumonia and COVID-19, 
variations in image collection conditions due to the risk of infection can lead to subtle differences in images. 
Existing deep learning models trained on such biased datasets may struggle to generalize effectively to data from different sources, 
resulting in overfitting issues.

The developed image translation model employs texture debiasing during the learning process, 
differentiating it from traditional models. Unlike standard models that might inadvertently alter content when adjusting textures, 
this model simultaneously considers error functions for both textures and contents. 
It extracts information on the input image's contents and textures from a different domain, 
enabling the generation of images with the texture of a different domain while retaining information about the original image's contents.

The model's superior performance was demonstrated on datasets with texture biases,
including those for distinguishing numbers, dogs and cats with various hair colors,
and COVID-19 from bacterial pneumonia. It outperformed existing methods on datasets with diverse biases, 
showcasing its effectiveness in tasks such as classifying multi-label numbers and distinguishing between photos, images, animations, and sketches.

Furthermore, the developed image translation technology extends to image manipulation, 
altering only the textures of an image while preserving its original contents. 
This method surpassed existing image manipulation techniques. 
The researchers believe that this technology will significantly enhance the robustness of AI models used in
industrial and medical fields where biased datasets are prevalent, 
making a substantial contribution to improving the performance and reliability of AI models in diverse environments for commercial purposes.