一种新颖的单幅图像超分辨率重建

A Novel Single Image Super Resolution Reconstruction

针对传统字典学习(dictionary learning,DL)的超分辨率重建方法,低分辨率图像块细节欠缺严重,导致重建的高分辨图像块无法获得最佳的高分辨率图像块。结合图像自身的自学习和自然图像库信息进行超分辨率重建。首先,通过自学习建立金字塔阶梯式训练图像集,结合协作表示(collaborative representation,CR)理论和支持向量回归(support vector regression,SVR)技术学习多层超分辨率映射模型,将得到的多张外部图像作为字典学习模型的低分辨率图像与原图训练高低字典对。图像重建时,先用自学习模型得到初步处理的X1,再用字典学习得到图像X2作为最终的重建图像。还提出了通过旋转和翻转图像训练分类的高低字典对,对实验结果有进一步的提升。实验结果证明,所使用的两个模型结合的算法,能够抑制传统方法不能解决的噪声问题,且与单纯的自学习模型相比,PSNR和SSIM也都有较为明显的提升。

Aiming at the super-resolution reconstruction method of traditional dictionary learning(DL), the lack of detail in low-resolution image blocks leads to the inability of reconstructed high-resolution image blocks to obtain the best high-resolution image blocks. This paper combines the self-learning of image itself and the information of natural image database for super-resolution reconstruction. Firstly, the pyramid step training image set was built through self-learning. Combining the theory of collaborative representation(CR) and support vector regression(SVR) technology, the multi-layer super-resolution mapping model was learned. The multi-external images were used as the low resolution dictionary learning model. The image was trained with the original dictionary. In image reconstruction, X1 was obtained based on self-learning model, and X2 was obtained through dictionary learning as the final reconstructed image. This paper also proposes a high-low dictionary pair classified by rotating and flipping image training, which further improves the experimental results. The experimental results show that the algorithm combined with the two models can suppress the noise problem which can not be solved by traditional methods, and PSNR and SSIM are also improved significantly compared with the simple self-learning model.