Image super-resolution using deep learning

Image super-resolution is a process to reconstruct a high-resolution image from a single low-resolution image. It has a wide range of applications in computer vision, medical imaging, satellite image processing, surveillance systems, etc. Recently, we have developed various deep learning models that significantly perform better than traditional methods. Our focus lies on improving perceptual quality and designing lightweight model architectures.

Efficient deep learning

Training and running deep learning models are usually expensive in terms of computation and memory due to their highly complex structures. This is a critical issue particularly when they are used in resource-constrained environments such as mobile and embedded systems. Therefore, it is highly desirable to make the models efficient without significant performance decrease. We are working on parallelized network training and model structure optimization.

Multimodal deep learning

Often, heterogeneous sensors are simultaneously involved, which produce multimodal data. For instance, human actions can be recorded by cameras, depth cameras, accelerometers, and gyroscopes at the same time. They are mutually complementary, thus learning from them is beneficial to maximize the performance. We have recently developed a deep learning architecture that can integrate multimodal data. It is not only effective in terms of performance, but also robust against partial absence of data and modalities.

Deep learning for graph data

Unlike audio, image, and video signals, some types of data are not well defined on the Euclidean space but need to be represented by graphs or graph signals, e.g., brain signals, social network data, sensor network data, traffic data, etc. Traditional deep neural networks specialized for audio or image signals cannot be used for those types of data. We work on developing new types of deep neural networks that can handle graphs and graph signals.

Adversarial attacks and defenses for deep learning models

Despite their good performance, deep learning models are known to be vulnerable to adversarial attacks appearing as small perturbation in the input data. We are working on this issue for various deep learning applications. First, we work on identifying vulnerability of deep learning models and understanding it theoretically and experimentally. Second, we work on developing defensive methods and robust model architectures against adversarial attacks.

Subjective and objective perceptual quality assessment

Measuring perceptual quality, known as quality of experience (QoE), of multimedia is fundamental for developing and evaluating various human-centric image and video processing algorithms. We are working on important issues around QoE, including developing efficient subjective quality evaluation methods, evaluating image/video compression methods, studying perceptual quality of 4K UHD, 3D, video scalability, benchmarking objective quality metrics, etc.

Brain-computer interface (BCI) for multimedia experience

BCI is a powerful way to understand users' perception directly by monitoring brain activities. We have been working on investigating effective BCI techniques using EEG and fNIRS for implict monitoring of users' perception of given multimedia content. Here, perceptual dimensions include satisfaction, quality, immersiveness, emotion, sense of reality, etc. In particular, we focus on understanding perception for novel types of media such as 3D, high dynamic range (HDR), 4K UHD, etc. Furthermore, peripheral physiological signals such as skin temperature, skin conductance, and heart rate are also considered, which have potential to complement brain activities.

Aesthetic quality of images and videos

Aesthetic quality of images and videos is important for human perception, but not easy to automatically measure. We work on understanding how humans perceive aesthetic quality of visual media, and how such understanding can be used for automatic image and video processing such as aesthetic quality enhancement. We have studied the notion of "subjectivity" indicating the fact that for some images and videos, aesthetic opinions of different people vary significantly. We have also developed automated video editing methods that can generate aesthetically improved versions of user-generated videos.

QoE-aware video communications

The demand for wireless video delivery is ever increasing with the spread of powerful portable devices. A key issue for successful video delivery services is how to satisfy users' quality of experience (QoE) for the delivered content. We have been working on developing techniques for QoE-optimized video communications and QoE measurement for video communications, particularly based on massive-MIMO and scalable video coding (SVC).

Visual attention analysis and modeling

The human visual attention allows to focus on a small region in the scene and ignore others at each time. Understanding how visual attention works via eye-tracking has importance in designing perceptually optimized multimedia processing and machine vision algorithms. Examples of our work on visual attention analysis include audio-visual focus of attention, gaze pattern analysis for temporal resolution changes, gaze pattern analysis for video packet loss, etc.

Music recommendation

Music streaming services are prevalent in these days. In such services, it is important to understand listeners' satisfaction and recommend relevant songs at the right moment. We are seeking answers to various research questions such as: how to generate a music playlist satisfying users, how to recommend particular music clips based on listeners' states, how to sense listeners' response to given songs, how to predict popularity of songs, etc.
Back to top