含噪图像的快速超分辨率重建算法

Fast super-resolution reconstruction algorithm for noisy image

现有算法大多假设输入图像是不含有噪声的。但与实际情况相反,在生活中获得的图像多数是含有噪声的。本文对含噪图像的超分辨率重建问题进行研究,并提出一种可以快速实现的算法。首先,借鉴传统算法中高、低分辨率字典的训练方法,在此基础上将低分辨率图像块的纹理结构加入字典的训练过程。值得注意的是,这里的低分辨率图像块和高分辨率图像块具有相同的图像尺寸,前者是通过双立方插值得到的。其次,由于字典训练过程中使用的实例图像是不含有噪声的,因此面对不同程度噪声的输入图像并不需要重新训练字典。在重建过程中,通过使用稀疏字典的列原子作为匹配对象从而大大降低了计算成本,并对输入的特征向量和稀疏字典做了归一化处理,提高了精度。根据输入的特征向量和匹配对象的相似程度选择k个相似块,并通过权重限制模型完成对相似块的权值分配,从而重构出对应的高分辨率图像块。最后,通过加权平均重建了原始估计的高分辨率图像和去噪后的低分辨率图像,再将两幅重建图像与迭代反投影相结合,得到最终估计的高分辨率图像。在自然图像上验证了本文算法,并与先前报道的算法进行比较,其结果优于其他算法并具有较好的鲁棒性。

Most existing algorithms assume that the input image does not contain noise. However, contrary to the actual situation, most of the images obtained in life contain noise, the super-resolution reconstruction of noisy image is studied, and a fast algorithm is proposed. First of all, based on the training methods of high and low resolution dictionaries in traditional algorithms, the texture structure of low resolution image blocks is added to the training process of dictionaries. On this basis, texture structure of low resolution image block is added into the training process of dictionary. It is worth noting that the low-resolution image block and the high-resolution image block here have the same image size, and the former is obtained by bicubic interpolation. In addition, because the example image used in dictionary training does not contain noise, the input image with different noise levels does not need to be retrained. In the reconstruction process, the column atom of sparse dictionary is used as the matching object, which greatly reduces the calculation cost, and the input eigenvector and sparse dictionary are normalized to improve the accuracy. According to the input eigenvector and the similarity degree of matching objects, k similar blocks are selected, and the weight distribution of similar blocks is completed by weight restriction model. Then two reconstructed images are combined with iterative back projection to get the final estimated high-resolution image. The algorithm in this paper is verified on the natural image, and compared with the previously reported algorithm, the result is better than other algorithms and it has better robustness.