以範例為基礎之影像解析度增強方法

Abstract

When taking a photograph using digital devices such as digital cameras, usually we are not able to perfectly duplicate the scene we want to capture due to the limits of camera devices and storage space. Instead, we can only to sample the scene and store the color information of discrete space locations in the form of image pixels. The above fact arose a problem when we want to display an image on a bigger display device or zoom-in the image for checking details: There are not enough information to display an image in any resolution higher than what it was taken originally. Similarly, If we scale down the resolution of an image due to reasons like storage constraints, we will not be able to scale it back easily. The super-resolution problem is a heavily ill-posed problem, which means that a perfect solution does not exist. Which means, it is impossible to ”enlarge” an image perfectly. However, this also results in an interesting and useful research subject: How can we produce a better enlargement result with only the limited information we have? In this thesis, we assume that the after the user took a picture of the scene (target image), he/she may also took one or more pictures of that scene from a closer position (reference image), or can obtain such images from other sources (like internet photo databases etc.). In order to enlarge the target image, we first adapt a modified general examplebased algorithm to enlarge the target while trying to reduce noises often seen in results of such algorithm. Then we match the target and reference images in order to find their relative positions. Since reference images are taken closer to the scene, they include more detail information. The detail information can be used to recover the missed details at the same location in the enlarged target image. Finally, we adapt a texture transfer algorithm to synthesize details for textures in the enlarged target image similar to those on the reference images. Our result is better than traditional interpolation methods not only in the areas covered by the reference images, but also uncovered areas because of the modified general example-based method. It is also a highly flexible method since the number of reference images required is not a fixed number.