基于局部自相似性和奇异值分解的超采样图像细节增强  

Detail Enhancement for Image Super-Resolution Algorithm Based on SVD and Local Self-Similarity

作　　者：肖进胜[1,2] 高威[1] 彭红[1] 唐路敏 易本顺[1] 

机构地区：[1]武汉大学电子信息学院,武汉430072 [2]武汉大学测绘遥感信息工程国家重点实验室,武汉430079

出　　处：《计算机学报》2016年第7期1393-1406,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(61471272);国家留学基金;广东省自然科学基金(2014A030310169)资助

摘　　要：基于局部自相似性的图像超分辨率算法中存在面块或线条现象,导致图像纹理不自然,细节信息丢失严重,针对这个问题,文中提出了一种基于局部自相似性和奇异值阈值化的细节增强图像超分辨率算法.在通过轮廓模板插值得到初始超分辨率图像的基础上,利用奇异值分解及阈值化去噪提高局部自相似性提取高频信息的准确性,解决伪高频噪声成分造成的面块或线条痕迹.该细节增强超分辨率算法主要分为4个部分,即初始插值、块匹配搜索、奇异值阈值、细节合成.首先选取Pascal的轮廓模板插值算法得到初始超分辨率图像,该算法在保持插值图像整体轮廓和细节等方面都有着很好的效果优势.块匹配搜索部分,文中算法由初始超分辨率图像和原始低分辨率图像分别提供参考块和学习块,在单一尺度上进行匹配,更多地利用了原始图像的信息,复杂度也更低,即利用原始低分辨率图像的局部自相似性提供先验知识,进行块匹配学习,找到初始超分辨率图像块在原始低分辨率图像块中的最佳匹配块,提取出最佳匹配块的高频信息;然后利用奇异值分解将高频信息矩阵分解到两个正交子空间中,并选取合适的阈值对奇异值矩阵进行软阈值化处理,剔除高频成分中能量较小的伪高频噪声成分,得到更为准确的高频细节信息.最后为得到最终超分辨率图像,有效地实现超分辨率图像的细节增强,利用有效奇异值对应的奇异值矢量重构高频图像块矩阵,并在初始超分辨率图像上进行细节合成.合成过程中,选择中心对称的高斯函数对图像块进行加窗处理,以抑制分块重叠带来的重叠区影响.实验结果数据表明,文中算法不但能明显解决由于伪高频成分导致的面块或线条现象,重建出的图像纹理细节更真实丰富,纹理结构和边缘特征的清晰度和对比度较高,得到的高分辨率图像视觉效果也�The facet or line phenomenon existed in the super-resolution algorithm based on local self-similarity leads to unnatural texture and serious loss of details.A detail enhancement super-resolution algorithm based on singular value threshold and local self-similarity is proposed here. On the basis of obtaining the initial SR image with the contour stencils interpolation,the denoising of Singular Value Decomposition （SVD）threshold is used to improve the accuracy of the highnbsp;frequency information extracted by Local Self-Similarity （LSS）.The facet or line phenomenon caused by the pseudo high frequency is suppressed too.The detail enhancement image SR algorithm comprises four major steps：initial interpolation,block matching,SVD and thresholding, detail synthesis.First,we have selected the contour stencils interpolation algorithm proposed by Pascal to obtain the initial SR image.The algorithm has a strong effect advantage in maintaining the whole contour and details of the high-resolution image.In the part of block matching,the algorithm in this paper directly uses the information of original LR image and initial SR image to provide the reference and learning blocks.The matching is done in single scale which utilizes more information of the original image and has lower computational complexity.That is to say, the local self-similarity of the original low resolution image is used as a priori knowledge in block matching,Then find the initial SR image patch’s most similar image patch in the original image. The high frequency information of the most similar image patch is extracted.Then the information matrix of high frequency block is decomposed to two sub-spaces by singular value decomposition and the singular value matrix is processed by the soft threshold with a suitable threshold.Thus the pseudo high frequency noise of small energy is removed by the soft threshold,which can get the more accurate high frequency detail information.Last,to obtain the final super-resolution image and achieve the detail enhan

关 键 词：超采样 细节增强 奇异值分解 局部自相似性 块匹配 计算机视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]