Estimating & improving image quality for human perception with generative and vision language models

Abstract

The widespread use of smartphones has made image-based communication an essential part of our daily lives. However, images are often degraded by noise during acquisition and by compression processes, leading to reduced perceptual quality and hindering effective communication.

This presentation introduces recent research on estimating the perceptual quality of degraded images and improving them using modern computational approaches. It focuses on how diffusion-based generative models and vision language models can be applied to enhance image quality from a human visual perspective.

Speaker

Takamichi Miyata, Professor, Chiba Institute of Technology

Takamichi Miyata received his Ph.D. in Engineering from Tokyo Institute of Technology in 2006. He served as an Assistant Professor at Tokyo Institute of Technology from 2006 to 2012, and as an Associate Professor at Chiba Institute of Technology from 2012 to 2015. He was appointed Professor at the Faculty of Engineering, Chiba Institute of Technology, in 2015, and has been serving as Professor at the Faculty of Advanced Engineering since April 2016.

His research focuses on mathematical optimization and deep learning, particularly on fundamental technologies for image processing and recognition using vision-language models. He has published 38 peer-reviewed journal papers—including in the IEEE Transactions on Image Processing (Impact Factor: 10.6)—and 48 international conference papers. His top three publications have been cited 137, 45, and 43 times, respectively. He received the IEICE Best Paper Award in 2014.