基于系数复用和字典训练的图像超分辨率算法

An Image Super-resolution Reconstruction Algorithm Based on Coefficient Reuse and Dictionary Training

在基于学习的图像超分辨率重建过程中,字典的选择和训练是其中的关键环节,但是传统的字典训练算法存在计算量大、训练速度慢等缺点,导致整个重建过程耗费时间长,重建图像在细节上表现较差,影响了其视觉效果与使用价值。针对上述字典训练中存在的问题,提出了一种改进的基于系数复用和字典训练的图像超分辨率算法。该算法对传统的K-SVD算法中的字典训练阶段进行了改进,利用信号的稀疏表示原理,同时结合正交匹配追踪中的系数复用算法,较好地解决了字典训练速度慢、重建图像质量低等问题。实验结果表明,与经典的双三次插值和改进前的K-SVD图像重建算法相比,该图像重建算法较好地复原了图像的高频细节信息,提高了重建图像质量,同时大幅度降低了字典训练时间。

In the reconstruction of super-resolution image based on learning,the selection and training of the dictionary is an important step.However,traditional dictionary training algorithms have many disadvantages,such as large amount of calculation,slowtraining speed and so on,leading to the long time-consuming of entire reconstruction process and the poor performance of detail for reconstructed image,which affect its visual effects and practical value.Aiming at the above problem in dictionary training,we propose an improved image super-resolution algorithm based on coefficient multiplexing and dictionary training.In this paper,we improve the dictionary training stage in the traditional K-SVD algorithm,and solve the problem of the slowtraining speed and the poor reconstructed image by means of the sparse representation of signal with the coefficient multiplexing algorithm in orthogonal matching pursuit.Experiments showthat compared to the classical bicubic interpolation and the traditional K-SVD image super-resolution algorithm,the proposed algorithm can recover the high-frequency details better,improve the quality of the reconstructed image,and substantially reduce the training time of dictionary.