多层次轮廓约束的图像放大算法

Image Magnification with Multi-Level Contour Constraints

为解决图像放大过程中有效地保证边缘锐化的图像插值难题,提出多层次轮廓约束的图像放大算法.首先利用检测算子对图像进行预处理,将图像分为边缘区域、平坦区域;其次,针对图像边缘区域进行自适应梯度扩散获取适当的边缘轮廓层作为图像放大约束;最后对轮廓层直接进行曲线插值重采样,不额外增加边缘层数,以保证放大后的图像在视觉上的边缘清晰.对于非轮廓层的平坦区域,构建双三次Coons插值曲面并进行重采样,保持了平坦区域的平滑性.测试图像为自然图像和医学图像,自然图像的来源是set5和set14测试集,实验对比方法主要从客观效果、视觉效果、时间复杂度3个方面进行比较.实验结果表明,采用该算法得到的放大图像不仅可以保持轮廓清晰,且PSNR及SSIM指标超过了大多数经典的插值算法以及目前流行的基于机器学习的算法.

Effective edge sharpening in image enlargement is always a difficult problem in image interpolation, and in order to solve this problem, an image enlargement algorithm with multi-level contour constraints is proposed. The detection operator is used to preprocess the image, and the image is divided into edge region and flat region. Adaptive gradient diffusion is applied to the edge region of the image to obtain an appropriate edge contour layer as an image magnification constraint. Finally, the contour layer is re-sampled directly through curve interpolation without adding additional edge layers to ensure that the enlarged image has clear visual edges. For the flat area of the non-contour layer, the bi-cubic Coons interpolation surface is constructed and resampled to maintain the smoothness of the flat region. The test images are natural images and medical images. The source of natural images is set5 and set14 test sets. The experimental comparison is mainly made from three aspects: objective effect, visual effect and time complexity. The experimental results show that the enlarged image obtained by the new algorithm can not only keep the contour clear, but also the PSNR and SSIM indexes exceed most classical interpolation algorithms and the popular machine learning-based algorithms.