挖掘理想重建图像自相似性的超分辨率

Image Super-resolution by Exploiting Self-similarity of Ideal Reconstruction

为了解决图像超分辨率过程中训练步骤对海量数据的过于依赖、先验泛化能力不强等问题,进一步提高重建图像的质量,提出了一种新的图像超分辨率算法.首先对图像自相似性理论进行扩展,指出理想重建图像自相似性表现极为强烈,而受降质因素干扰的重建图像自相似性则会明显减弱.本文将这一规律视为先验,通过构建联合高斯混合模型对其进行描述,这使得每个重建图像片的自相似性都能够用一个特定的高斯分布进行刻画,最后算法以迭代的方式分片重建整幅高分辨率图像.在为每个高分辨率图像片建模的过程中,为了使训练样本具有较强的一致性,仅使用输入图像中与其空间位置相近的图像片进行训练.该算法避开了易于引入误差的最近邻域查找步骤,且成本函数存在解析解.实验表明该算法重建图像清晰、自然,重建结果中的显著边缘和纹理结构都得到了有效保持,正确的高频信息得到了明显恢复.在将BSD500部分数据集放大3倍的实验中,本文算法的PSNR平均值高于MMPM算法0.529 db,SSIM平均值高于MMPM算法0.030.

To solve the problems such as over-reliance on massive data and weak prior generalization ability in the training procedure of image super-resolution,thus further to improve the quality of reconstructed high resolution image,a new image super-resolution algorithm was proposed.This paper firstly extends the theory of image self-similarity and points out that the self-similarity of ideal reconstruction image is extremely strong,but this property can be sharply weakened when the reconstructed image is attacked with some degradation factors.Then this discovery is considered as a prior and described by constructing a joint Gaussian mixture model,so that the self-similarity of each reconstructed image patch in the prior term can be represented by a specific Gaussian distribution.For maintaining the training samples'consistency,only the image patches extracted in the input image closed to its spatial position are permitted to join in the modeling process for each high-resolution image patch.This style can avoid the step of finding the nearest neighbors which is liable to introduce errors.Finally,the whole high-resolution image can be reconstructed patch-wise in an iterative way.Extensive experiments demonstrate that the reconstructed images generated by the proposed algorithm are clear and natural,in which the salient edges and texture structures are effectively preserved,and the correct high-frequency information is recovered.The 3×super-resolution experiment in BSD500 shows that the average PSNR is higher 0.529 db than the state-of-the-art algorithm MMPM,and the average SSIM is 0.030 higher than MMPM.