基于自学习的稀疏正则化图像超分辨率方法

如何设计既能够保持边缘与纹理结构又具有较低计算复杂度的图像超分辨率算法是目前该领域有待解决的难点问题。在Bayesian统计框架下建立了一种新的基于稀疏正则化的图像超分辨模型。模型中的保真项度量理想图像在退化模型下与观测图像的一致性,稀疏正则项刻画理想图像在词典下的稀疏表示。该模型还引入了图像的非局部自相似性和超拉普拉斯先验作为正则化约束。为使稀疏域更好地表征高分辨率图像,选取高分辨率图像块的高频特征进行稀疏表示,由此增强了稀疏模型的有效性。将词典学习融入到超分辨率重建过程中,即直接从当前估计的高分辨率图像特征块学习词典,与从训练样本库中学习词典相比,这种自学习的方法对不同图像的自适应性更强,并且减少了运算量。实验结果表明,该方法可以重建清晰的图像边缘,减小振铃效应,并且对噪声具有很好的鲁棒性。

Self-learning image super-rsolution method based on sparse representation

It is difficult to design an image super-resolution algorithm that can not only preserve image edges and texture structure but also keep lower computational complexity. A new super-resolution model based on sparsity regularization in Bayesian framework is presented. The fidelity term restricts the underlying image to be consistent with the observation image in terms of the image degradation model. The sparsity regularization term constraints the underlying image with a sparse representation in a proper dictionary. We also introduce the non-local self-similarity and hyper-laplacian prior as regularization constraints into the model. In order to make the sparse domain better represent the underlying image, high-frequency features extracted from the underlying image patches are used for sparse representation, which increases the effectiveness of sparse modeling. The dictionary learning into the super-resolution process is incorporated, in this way the dictionary can be learned directly from the currently estimated high-resolution image feature patches. Compared with learning dictionary from pre-collected training data, the self-learning method has stronger adaptability to different images, and reduces computation cost. Experimental results show that our method can reconstruct clear and sharp image edges, reduces ringing effects and has good robustness to noise.