基于轮廓模板和自学习的图像纹理增强超采样算法

发布时间：2018-04-19 06:11

本文选题：超采样 + 图像插值　；参考：《自动化学报》2016年08期

【摘要】：提出一种以轮廓模板插值和局部自学习相结合的图像纹理增强超采样算法,有效地恢复了插值图像丢失的细节纹理,抑制了插值图像边缘的扩散.该方法通过局部自相似性在原始低分辨图像中估计高频信息,对轮廓模板插值图像的细节纹理进行了恢复.其中,为了弥补轮廓模板插值缺少先验知识的缺陷,将原始低分辨率图像的高频信息作为先验知识.为了保证估计的高频信息最优,匹配的过程中采用双匹配,相比较于全局搜索和小窗搜索,提高了效率并保证了匹配精度.此外,使用高斯模糊代替了传统提取高频信息的方法,简化了算法的复杂度,提高了准确性和效率.对估计得到的高频信息采用高斯函数加窗,以减小估计出错和重叠区的混叠影响.本文算法的训练库由原始低分辨图像自身和插值图像构成,节省了生成训练库所需的时间和空间.训练库的简化使得高频信息的估计可以多尺度进行,算法效率得到进一步优化.理论分析和实验结果表明,相比传统的基于插值、基于自学习的图像超分辨率方法,本文方法获得更好的实验结果,主观效果得到明显改善,有效地恢复了图像的纹理细节,提高了图像边缘锐度,避免了产生锯齿等人工效应,客观指标得到提高.
[Abstract]:An image texture enhancement oversampling algorithm based on contour template interpolation and local self-learning is proposed, which can effectively restore the lost detail texture of the interpolated image and restrain the edge diffusion of the interpolated image.In this method, the local self-similarity is used to estimate the high-frequency information in the original low-resolution image, and the detailed texture of the contour template interpolation image is restored.In order to make up for the lack of prior knowledge in contour template interpolation, the high frequency information of the original low resolution image is regarded as prior knowledge.In order to ensure the optimal estimation of high frequency information, double matching is used in the matching process. Compared with global search and small window search, the efficiency is improved and the matching accuracy is ensured.In addition, Gao Si fuzziness replaces the traditional method of extracting high frequency information, which simplifies the complexity of the algorithm and improves the accuracy and efficiency.Gao Si function is used to window the estimated high frequency information in order to reduce the aliasing effect of the estimation error and overlapping region.The training library of this algorithm is composed of the original low-resolution image itself and the interpolated image, which saves the time and space needed to generate the training library.With the simplification of the training library, the estimation of high frequency information can be carried out on multiple scales, and the efficiency of the algorithm is further optimized.The theoretical analysis and experimental results show that compared with the traditional image super-resolution method based on interpolation and self-learning, the proposed method achieves better experimental results, and the subjective effect is obviously improved, and the texture details of the image are recovered effectively.The edge sharpness of the image is improved, the manual effect such as sawtooth is avoided, and the objective index is improved.
【作者单位】：武汉大学电子信息学院;武汉大学测绘遥感信息工程国家重点实验室;
【基金】：国家自然科学基金(61471272)资助~~
【分类号】：TP391.41

【相似文献】