Best Paper Award

Super Resolution Image Synthesis from a Light Field Camera Considering Its Physical Pixel Arrangement

Yuki UCHIDA, Kazuki OHASHI, Keita TAKAHASHI, Toshiaki FUJII

[Trans. Inf. & Syst. (JPN Edition), Sep. 2016]

  There is considerable interest in light field-related applications such as digital refocusing, depth estimation, free viewpoint video generation, etc. Such applications are based on acquisition of 3D scene information by means of a single device.
   To date, several methods have been studied in the light field image processing area, including multiview image generation from a raw image obtained by a light field camera, and a super-resolution technique for generating multiview images with higher resolution. There are three areas of difficulty in handling raw images. First, each pixel on an image sensor has only one color value among R, G, or B. Second, integer coordinates in a multiview image correspond to floating-point coordinates in the raw image. Finally, the micro-lens-array is aligned in a hexagonal layout. Generally, to overcome these problems, demosaicing, smooth interpolation, and resampling are utilized, although further problems are introduced.
   In this paper, the authors propose a multiview image reproduction method based on maintaining the pixel layout on the raw image. It is a very sophisticated approach, holding that avoiding interpolation processing, which causes mixing of different viewpoints, will solve the problem, in principle. Furthermore, utilization of ADMM, which is one of the alternative optimizations in convex optimization methods, is also effective for two related targets, which are depth estimation and super resolution.
   The proposed method offers high universality through a generalized approach although the evaluation experiments have only been conducted with a light field camera called eLytro Illum.f Since both the problem statements and their solution narrative are theoretically constructed, effectiveness of the proposed method would seem to be very high due to its high rationality. Furthermore, both the logicality and the readability are also high.
   From the above considerations, the proposed method, which originally was an integration of pixel-coordinate conversion process on a raw image into an observed model, can reproduce clearer multiview images with higher resolution than can be achieved with the conventional method. The proposed method is a fundamental and essential technique in light field studies, where future developments are expected. Consequently, this paper is highly deserving of a Best Paper Award.
Close