支持向量机多特征分类学习的超分辨率复原

Super-resolution Restoration Algorithm Based on SVM Multi-figure Classification Learning

支持向量机(SVM)单一特征分类学习的超分辨率复原算法通过离线建立分类模型和减少样本库规模,降低了传统基于范例学习算法的样本块误匹配情况,增强了图像质量和计算速度.但由于图像特征的多样性,此类算法易造成复原结果的不稳定.本文给出一种以支持向量机多特征分类学习为基础的复原算法,将图像对应的颜色和纹理分类信息存储在样本库中,经过预分类筛选出样本子集,在高频预测时段直接从多特征相似的样本子集里实施准确的匹配检索.实验结果表明,相比于传统算法,本文算法的PSNR和SSIM值均有了一定提升,进一步精确匹配了低分辨率图像样本库,提高了复原效果.

The SVM pre-classified super-resolution algorithm is based on single image feature and builds off-line disaggregated models.It reduces the mis-matching of tranditional example-based restoration algorithms,improves the image quality and running speed.However,the SVM-based algorithm easily leads to unstable recovered results because of the diversity of image features.For such problems,we propose a super-resolution restoration algorithm based on multi-figure classification learning.The algorithm saves the corresponding color-texture information into the sample set and selects object subset by SVM pre-classified learning.Then in the high frequency prediction process it makes precise matching search from the subset of sample database which has similar color and texture features with the object image.Experimental results show that compared with traditional algorithms,PSNR and SSIM are improved respectively.In addition,the proposed algorithm further reduces the matching range of low resolution image blocks and promotes the restoration effectiveness.