Abstract: This paper presents a new learning based super-resolution of image by introducing the Nonsubsampled Contourlet Transform（NSCT） and improved BP neural network.NSCT can recover the detail information better, as the improved BP can simulate the highly nonlinear, which is fast convergence and accuracy. For both super-resolution image and low-resolution image, it extracts the Contourlet coefficients of each sub-band training the improved BP network, then using the stable and restraining network realizes super-resolution reconstruction of image. The results show that this method is able to preserve the details of original image better and reduce the complexity of the network reconstruction at the same time, raise the accuracy, get significantly improved in reconstruction results.

Key words: super-resolution reconstruction, nonsubsampled Contourlet transform, improved BP neural network

摘要： 鉴于非下采样Contourlet变换（NSCT）系数包含原始图像各方向的所有细节信息，以及改进BP神经网络高度非线性映射的快速收敛和准确性，提出一种应用NSCT和改进BP神经网络的超分辨率图像重建算法。分别提取模拟超分辨率图像与相应低分辨率图像各方向子带的NSCT系数进行BP神经网络高度非线性映射训练，直至稳定收敛，并利用该网络实现超分辨率图像重建。实验结果表明该算法在很好保留图像细节的同时极大地降低网络重建复杂度，提高了重建的准确率，重建效果得到明显改进。

关键词: 超分辨率重建, 非下采样Contourlet变换, 改进BP神经网络