基于BP神经网络的超分辨率图像重建

Reconstruction of super-resolution image based on BP neural network

针对卫星图像成像过程中成像装置存在极限,导致图像分辨率低的问题,提出了基于神经网络的图像超分辨率重建(neural networks super-resolution reconstruction,NNSR)方法。该方法利用误差反向传播神经网络(back propagation neural networks,BPNN)对样本图像进行学习和训练,利用图像退化模型获取学习样本,采用向量映射加速BP神经网络的收敛,充分融合了低分辨率序列图像中的冗余信息。通过对训练好的神经网络分别进行样本仿真实验和泛化实验,验证了这种图像超分辨率重建方法的有效性。

The reconstruction of the super-resolution image based on neural networks （NN） is proposed to resolve the problem of image low spatial resolution because of the limitation of imaging devices. An error backpropagation （BP） algorithm is used to learn and train sample images in order to combine the redundancy information of low spatial resolution images sequences. Learning samples are acquired according to the image observation model. Vector mapping is established to speed up the convergence of NN. Simulation and generalization tests are carried on the well-trained NN respectively, and the reconstruction results with higher spatial resolution images verify the effectiveness and validity of BPNN based on vector mapping in the reconstruction of the super-resolution image.