Kai Cui, M.Sc.
Chair of Media Technology (Prof. Steinbach)
- Phone: +49 (89) 289 - 25823
- Room: 0504.EG.409
Kai Cui received the B.Sc. degree in electronic information engineering from Harbin Institute of Technology, China, in 2013 and the corresponding M.Sc. degree in 2015. He is currently working toward the Ph.D. degree at Technical University of Munich (TUM). He joined the Chair of Media Technology at TUM in 2015, where he is a member of the Research and Teaching Staff. His research interests include image and video processing, image/video coding and machine learning.
With the success of convolutional neural networks (CNN) in image processing, a number of CNN based algorithms have been proposed and many of them can outperform conventional approaches. However, most of them are focusing on high-level image processing tasks, e.g. image classification, object detection and recognition. For low-level image and video processing tasks, e.g. color image demosaicking, denoising, compression artifacts removal, image and video enhancement, they can also benefit from CNN based approaches. Conventional approaches are mostly based handcrafted priors, which may lead to other unwanted artifacts in some cases. With CNN and data-driven training process, usually better features can be extracted from the data and the model obtained will be more adaptive to the data. Here, we take color image demosaicking and compressed image quality enhancement as examples.
- Color Image Demosaicking
Color demosaicking (CDM) is a critical first step for the acquisition of high-quality RGB images with single chip cameras. Conventional CDM approaches are mostly based on interpolation schemes and hand-crafted image priors, which result in unpleasant visual artifacts in some cases. Motivated by the special characteristics of inter-channel correlations (higher correlations for R/G and G/B channels than that for R/B), we designed a 3-stage CNN structure for CDM which is shown in Fig. 1. In the first stage, the G channel is reconstructed independently. Then, by using the reconstructed G channel as guidance, the R and B channels are recovered in the second stage. Finally, high-quality RGB color images are reconstructed in the third stage.
- Compressed image quality enhancement
Lossy compressed images usually suffer from unpleasant artifacts, especially when the bitrate is low. In order to improve the image quality without spending extra bit-rate, decoder side quality enhancement becomes necessary. Most existing approaches focus on spatial information exploration, in which the quality enhancement is usually only performed on the luminance component or the gray-scale images which makes the inter-channel correlation is neglected. Motivated by the characteristics of compressed Images and the color demosaicking network we designed before, we designed a similar network for decoder side image quality enhancement, which is shown in Fig. 2. This structure can make the most of inter-channel correlation and is compatible with any existing image/video coding standards, we test several coding standards and the results show that our scheme can achieve noticeable objective and subjective quality improvement.