Page 24 - profile-ok
P. 24
研究群 | Research Laboratories
Multimedia Technologies Laboratory
Research Faculty Group Profile
Hong-Yuan Mark Liao Multimedia research covers a broad scope of techniques and rich applications, including works on 3. Lighting Normalization gle-pixel camera, compressive sensing (CS) has been able to directly
Research Fellow text, music, video, images, and 3-D animation. Its rapid progress has now become one of the piv- capture compressed image data efficiently. The compressed image
Electrical Engineering , Northwestern University otal factors affecting our daily life. Together with Biotechnology and Nanotechnology, Multimedia Lighting variation has long been a fundamentally important problem can be reconstructed using some CS reconstruction algorithms at
Chu-Song Chen is considered as one of the three most promising industries of the twenty-first century. in video content understanding. To deal with this problem, previous the decoder if the image has sparse representation (compressible) in
research includes training-based approaches that employ a large col-
Research Fellow The Multimedia group currently focuses its research efforts on two main areas: 1) multimedia lection of images for finding invariance, and decomposition-based some domain (e.g., DCT or DWT). By integrating the respective char-
Computer Science and Information Engineering , acteristics of DVC and CS, we will investigate a distributed compres-
National Taiwan University signal processing, and 2) multimedia applications. For the past decade, we have already accom- approaches that employ the intrinsic properties of images for de- sive video sensing (DCVS) framework to directly capture compressed
plished several significant results and developed various effective techniques for image process- lighting. In the past, we have developed a training-based approach
Wen-Liang Hwang ing, computer vision, computer graphics, video processing, multimedia security, and machine for moving cast shadow removal. However, this method requires a video data efficiently, where almost all computation burdens can be
Research Fellow learning. Some of these results have also been successfully introduced into their related industries period of training time for finding stable features. In the future, we shifted to the decoder, resulting in a very low-complexity encoder.
Computer Science , New York University
via technology transfers. In the next few years, we will continue to devote our research efforts will investigate variational approaches for image decomposition. By
Tyng-Luh Liu in advancing key fields in multimedia, including video forensics, video content analysis and un- separating an image into different types of signals, we seek to find 7. Game-theoretical Approach on Computing and Com-
Research Fellow derstanding, lighting normalization, common pattern discovery, multiclass object recognition, more stable features for variable lighting conditions. munication Resource Allocations
Computer Science , New York University distributed compressive video sensing, game theory for computing resource allocation, image
Chun-Shien Lu de-noising and de-blurring, and confocal microscope image stitching. We describe in detail some Because of the rapid development of Internet techniques that al-
Associate Research Fellow of these fields, below. 4. Common Pattern Discovery low many computational tasks to be handled by an Internet-based
Electrical Engineering , National Cheng-Kung infrastructure “in a cloud,” most future computing functions are being
University We will study the problem of finding common patterns in multiple shifted away from actual end users devices. Meanwhile, the evolu-
1.Video Forensics images or a sequence of images. In this area, we seek to find impor- tion of embedded and ubiquitous computing technologies means
tant common features based on un-supervised learning. Unlike exist- that personal computing devices will become smaller and smarter.
Video forensics is one of the most attractive research fields in recent years. Due to the rich ing methods that usually deal with only a single common pattern, In other words, cloud computing will take over much of the com-
amount of information that a video camcorder can record, police departments frequently use we will investigate a more general problem, when there are multiple putational load of personal devices, but the devices will need to
video when performing crime scene investigations. Collecting evidence from video is very common patterns or even no common patterns occurring within the do more in terms of computation capability. One can envision that
different from collecting evidence with simple images. The former is much more difficult than images. a tremendous number of users will compete for Internet resources
the latter, because it has to deal with spatial-temporal problems. A number of critical issues and, as a consequence, generate many conflicts of interest (called re-
that one may need to face include: low-resolution problems, motion across temporal axis, and source conflicts). This problem raises several challenges for the signal
illumination changes across consecutive frames. We plan to devote part of our research ef- 5. Multi-class Object Categorization processing community, as new technologies must be developed to
forts on video forensics in the next few years. The issues we would like to address include: resolve such conflicts. To this end, we propose using a game-theoret-
(1) video enhancing, and (2) video authentication. The sub-topics related to video enhancing One major obstacle hindering advances in developing object rec- ical-based approach.
are: motion de-blurring, and video stabilization and illumination stabilization across temporal ognition techniques has to do with the large intraclass feature varia-
axis. For video authentication, we shall deal with the issues of video inpainting and video copy tions caused by issues such as ambiguities from clutter background,
detection. various poses, different lighting conditions, possible occlusion, etc. 8. Image Noise Reduction and Blur Removal Algorithms
Another difficulty in addressing object recognition is that its current
application often deals with a large number of categories. While de- Reducing noise and removing blurring from images are important
2. Video Content Analysis and Understanding signing more robust visual features and their corresponding similarity fundamental problems in image processing. Although there are
measures has progressed significantly, the general conclusion is that many methods to recover noisy and blurred images, the solutions are
Video content analysis and understanding is also a very hot research area, with significant in- no single feature is sufficient for handling diverse objects of broad still not fully satisfying. For the noise reduction problem, could we
vestments from major technology players such as Google, Microsoft, yahoo, and IBM. In the categories. Taking into account these foregoing considerations, we propose a new model to obtain a higher PSNR denoised image? For
next couple of years, we will focus our emphasis on several basic issues. For video content aim to establish a general framework for addressing object recogni- the issue of de-blurring, because cameras are becoming smaller and
analysis, we will cover issues related to spatial-temporal content extraction and analysis, and tion over large and broad categories. In our previous research efforts lighter the solution to the problem is even more challenging. We are
heterogeneous features extraction and fusion for compact video representation. As for video on object recognition, we introduced a local learning approach to interested in solving these problems by proposing novel ideas.
retrieval, we will address issues related to the design of efficient representation schemes and designing ensemble kernel machines with proper localization and
the design of valid metrics for performing video retrieval. regularization. Our use of ensemble kernels has been shown to be
an effective way of fusing various informative kernels resulting from 9. Confocal Microscope Image Mosaicing
assorted visual features and distance functions, and has a marked
impact on succeeding related approaches. However, the proposed Confocal microscopes have been widely used in biological imaging
technique is still far from satisfactory. Our research efforts will con- because of the recent development of fluorescent probes and high-
tinue to better address a number of key issues, and consider informa- resolution imaging techniques. Biologists now depend more and
tion fusion in a broader sense (not just at the feature level), in order more on two or three dimensional images to visualize subcellular
to more effectively tackle the high complexity of multi-class object components in vivo. Because the very large size of microscope data
categorization. requires post-processing for their interpretation, this creates new
challenges for today’s image processing techniques. Due to the fact
that the optical resolution of a microscope is often limited, especially
6. Distributed Compressive Video Sensing in the axial (i.e. z), and fluorescence microscopy can image living cells
that move over time, the registration of two 2D image stacks become
Current low-complexity video codecs (including DVC) are usually de- very difficult and is almost never perfect. We are currently studying
signed to reduce encoding complexity to the order of that for still solutions for this problem.
image/intraframe video encoding. Recently, with the advent of a sin-
24 25