Domestic researchers have developed artificial intelligence (AI) capable of capturing detailed changes in images. It is expected to be applicable in image analysis, which requires high accuracy, such as in autonomous driving and medical image analysis.
A research team led by Professor Kim Jun-mo of the Korea Advanced Institute of Science and Technology (KAIST) announced on the 13th that they developed AI technology to understand image changes by mimicking human cognitive processes.
Computer vision is a technology that mimics human visual functions to enable AI to evaluate and predict images independently. It primarily trains on data generated by transforming images. However, in this case, AI can miss detailed changes, which decreases its performance. This means its applicability in fields like autonomous driving, which requires an understanding of subtle changes in the road, is limited.
The research team developed 'STL,' specialized in enabling AI to learn transformations of images independently. This solution addresses the issue of AI misjudging images with different transformations as similar while enhancing its sensitivity to distinguish between original and changed images. It is akin to a person playing a hidden object game, where the AI learns to find the altered parts within the images.
STL has shown to be approximately 42% more effective compared to existing AI systems. In experiments to differentiate images and identify transformed sections, it demonstrated a lower error rate than conventional AI. It also displayed superior ability in assessing relationships between transformed images and originals and distinguishing the types of transformations applied.
Professor Kim Jun-mo noted, 'STL has demonstrated the ability to learn complex transformation patterns and effectively reflect them in expression spaces,' adding, 'It will play a key role in various applications as it can learn transformation information without human input.'
The research findings were presented at the international academic conference NeurIPS 2024 on the 11th.