Super-resolution reconstruction based on image characteristics

Abstract

To increase the image resolution, various approaches have been proposed over past several decades. Among them, super-resolution has recently been actively researched. Super-resolution is a process of combining multiple low-resolution images to produce a higher resolution image. In conventional super-resolution algorithms, a single scheme is usually used for a whole image, regardless of region characteristics. Since an image consists of various regions having different characteristics, these algorithms may not provide equally good performance for all regions. In order to alleviate this fundamental drawback of the conventional approach, this dissertation presents a region-based super-resolution algorithm. In the algorithm, an image is first analyzed and segmented into smooth, intermediate, and edge regions. According to the region type, a different regularization term and different reconstruction parameters are used. The reconstruction parameters are independently managed and optimized depending on the region type so as to maximize the visual quality of the whole image. This dissertation also deals with regularization for the region-based super-resolution. Since super resolution is an ill-posed problem, considering regularization as a means for picking a stable solution is very useful. One of the most widely referenced regularization cost functions is the Tikhonov cost function using the Laplacian operator. The operator tends to reduce noise but to blur edges due to its omni-directional characteristic. In order to restore the continuity and sharpness of edges, directional smoothing is more appropriate than omni-directional smoothing. Therefore, we propose an adaptive diffusion regularization scheme which uses isotropic or anisotropic diffusion to efficiently reduce undesired noise and restore edges. Experimental results show that the proposed region-based super-resolution and the proposed adaptive diffusion regularization improve the objective quality as...