基于识别的单向多样学习图像超分辨率算法设计与实现

Design and Implementation of One-Pass Manifold Learning Image Superresolution Algorithm Based on Recognition

提出了一种新的图像超分辨率处理算法。首先建立训练图像集,然后对待处理图像和训练集中的特征图像对进行分割、光栅排列和对比度正则化等适当的预处理。待处理图像上的每个局部图像块在训练集中进行多样学习,以获得低分辨图像上不同区域缺乏的高频细节信息,最后使用这些信息预测生成超分辨率图像。实验结果表明,文章算法得到的高分辨率图像能够较大程度上提高图像质量。

A new image superresolution processing algorithm is proposed. Firstly, a training image set is constructed. Secondly, some preprocessing operations are adopted for the object image and the character image of training image set, such as image segmentation, raster permutation, and regularization of contrast. Each local block of the object image suffers the manifold learning from the training set, thus some high frequency detail information, which is scarce for the low-resolution image, can be obtained, finally, the high frequency detail information can be used to predict and produce the superresolution image. Experiment results show that the proposed algorithm is able to improve the image quality comparatively obviously.