Research

My research interest is to explore the possibilities in the acquisition, manipulation and exhibition of visual data. Up to now, this interest has driven me to three particular topics: 1) to extend the boundary of image processing with online photo collections, 2) to capture informative images by redesigning the principle of photography and 3) to give a sense of intelligence to traditional image processing techniques. For the first one, I named it "Collective Image Processing" in view of its joint nature as described below.

Thanks to the maturity and popularity of the digital camera technology and the Internet, online photo collections now grow at a rapid speed both in the size and the number. This new type of data provides unprecedented opportunities, while posing great challenges for conventional vision algorithms to overcome. Seeing the rising importance, I hope to break the existing limits in image processing by utilizing the rich information lying in photo collections. In particular, I want to transfer the necessary data from other images, so that each of our editing operations will be supported by "ground truths" which are not originally available. With this additional information, we can prevent false decisions which are often inevitable due to many problems' ill-posed nature. This will bring the traditional task of image processing to a brand new era. We are currently approaching the goal by applying image reconstruction methods to combine data from different images. For this purpose, we* have got fundamenal results in the case of single-scene online photo collection.

*: Our team now includes Cheng-Yi Chiang, Chia-Kai Liang and Professor Homer Chen.

Computational photography is a hot research topic which attempts to overcome the limitation of traditional photography through computational methods. I had been working on a high quality light field camera based on aperture multiplexing. Our system captures a high angular resolution and full spatial resolution light field, which is never achieved before by a single camera. The above images show the dense depth maps (mouse rollover) extracted from captured light fields using my multi-view stereo algorithm. The estimated depth maps could then be used to perform digital refocusing and view interpolation. In the future, I would like to design improved acquisition scheme to loosen the static scene restriction inherent in multiplexing. I am also looking into other design possibilities for informative visual acquisition.

Traditional image processing tries to apply post-processing on captured images for better visual experience. Fascinated by the effect of Seam Carving, I am currently sketching a resizing algorithm which could integrate image retargeting with texture synthesis. Similar concepts were recently proposed by "Inverse Texture Synthesis" and "Bidirectional Image Summarization", but they are either restricted to specific types of images or producing blurry and unstable results. These methods also suffer from the time-consuming backward nearest neighbor search which makes their running time ranges from minutes to even hours. To deal with these problems, I am trying to model the visual completeness in a more computationally succinct manner so that we could totally avoid the neccesity of backward search.

Tai-Hsu Lin

Research interests



We** applied dimension reduction for discrete minimization of the bidirectional similarity energy on general images. The color chart shows the locations of copied source pixels in the original image. For more information, please follow the link.