关键词: 图像超分辨, 支持向量回归（SVR）, 核非线性回归（KNR）

Abstract: Learning-based superresolution algorithm is one of the most potential techniques in image processing area in recent years. With this type of methods, a superresolution image may be properly restored by learning a certain model from examples. However, it requires a large quantity of samples which precludes its use in most practical situations. To tackle the problem, it tries to establish the local mapping models between a low resolution image and its interpolated version with required resolution, by dividing the two images into blocks and grouping them. For each group, a mapping model from a high resolution block values to a low resolution pixel value is established using Support Vector Regression（SVR） or Kernel Nonlinear Regression（KNR）. Interpolated pixels are predicted again using these models, and thus better superresolution result is obtained. The effectiveness of the proposed algorithm for single image based superresolution is illustrated by some experimental results.

Key words: image superresolution, Support Vector Regression（SVR）, Kernel Nonlinear Regression（KNR）