基于分组SVR和KNR的单帧图像超分辨

Single-frame image superresolution using grouping SVR and KNR

基于学习的图像超分辨是超分辨领域的一类新方法,该方法通过建立映射模型有针对性地对图像目标进行恢复,取得较好的超分辨效果,但往往需要大量学习样本,实际情况中一般难以满足。在无高分辨清晰图像库作为训练样本的前提下,从低分辨图像与其插值图像之间的关系出发,引入分组的思想,采用支持向量回归(SVR)或核非线性回归(KNR)对"组"建立局部映射模型,利用局部模型针对性地重新估计被插值的像素点。结果表明该方法有明显的超分辨效果。

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.