为了能够完成噪声图像的超分辨率重构, 提出了一种基于稀疏表示的噪声图像超分辨率重构方法, 可以同时完成图像去噪和超分辨率重构。首先, 对样本图像和低分辨率图像进行块划分, 建立样本库。其次, 建立图像退化模型, 采用相似样本加权平均的方式对输出的高分辨率图像块进行表示。根据输入的低分辨率图像块, 计算样本块与输出的高分辨率图像块之间的相似性。提出了一种相似性描述方法, 能够很好地解决噪声带来的影响。然后, 采用相似性对稀疏编码优化模型进行惩罚, 提出一种权值求解模型。模型可以自适应的搜索相似样本块而不需要预先设定相似块的个数。最后, 求解权值, 根据权值和样本块重构高分辨率图像块, 并重构高分辨率图像。实验结果表明: 所提出的方法较其它常见超分辨率算法的峰值信噪比可提高0.5dB左右, 重构的图像细节更丰富, 去噪效果更好, 更适合实际应用。

In order to complete the super-resolution reconstruction of noise images, a reconstruction method of noise images was introduced based on sparse representation,which could complete image de-noising and super resolution reconstruction simultaneously. Firstly, block size division was made for sample images and low-resolution images and the sample database was established. Secondly, the image degradation model was built and the way of weighted average was used for similar samples to represent the output image block with high resolution. Then, according to the input low-resolution image block, the similarity between sample block and output high-resolution image block was calculated. In addition, a similarity description method which could better reduce the influence bought by noises was proposed. Using the similarity to punish the sparse coding optimization models, a weight solving model was established . And the similar sample model could be self-adaptively searched by the model rather than being set the number of similar blocks in advance. Finally, the image block with high resolution as well as high-resolution images were reconstructed, according to the solved weight and input sample block . The result of experiment shows: compared with the other common super resolution algorithms, the peak signal to noise ratio of the mentioned method improves approximately 0.5 dB; and the reconstructed image with more details has better de-noise effect and is more suitable to practical use.