Speaker: Thomas Kerndlmaier
With the rapid development of convolutional neural network (CNN), CNN has been adopted in many compressed image enhancement approaches and achieve the state-of-the-art performance. Existing approaches need to be trained separately for a specific quality factor or a limited range of quality factors. They perform not well when the image quality level is mismatching with the trained image or it never appears in the training dataset.
In this thesis, we investigate the potential to use a single trained CNN model to enhance the images compressed with any quality factors. We prove that the a single trained CNN successfully handles the quality independent enahncement task with apropriate training strategy, and the performance compromise is minimal compared to a separate model for each quality level. This makes the training more manageable and the CNN approach more practical in a real image codec. Mr. Thomas Kerndlmaier will present the work he performed and the results obtained during his master thesis.