基于径向基函数神经网络的超分辨率图像重建

为了突破成像极限,经济可行地获取高质量的卫星图像,提出了一种基于径向基神经网络的超分辨率图像重建算法。以径向基神经网络为基础,依据卫星图像退化模型获取网络训练所需的学习样本图像,采用向量映射的方式加速网络收敛。其中,径向基函数的中心、宽度及网络的隐含层数、连接权值是决定径向基神经网络的关键参数,直接关系到网络的重建性能。采用最近邻聚类算法,动态地建立起基函数的中心及宽度,自适应地确定网络的隐含层数及连接权值。建立起的径向基函数神经网络显著地提高了图像重建性能和网络收敛速度(221s即可收敛)。仿真实验和泛化实验表明,训练好的径向基神经网络可以有效地进行卫星图像的超分辨率重建,效率高,误差小。

Super-resolution image reconstruction based on RBF neural network

In order to break through the limitations of imaging devices and to resolve the problems of Super-Resolution Reconstruction (SRR) of a satellite image, an image reconstruction based on the Radial Basis Function Neural Network (RBFNN) is proposed. First, learning sample images are acquired according to a satellite image observation model and the vector mapping is established to speed up the convergence of RBFNN. Then,the nearest neighbor clustering algorithm is used to dynamically establish the centers and widths of RBF, and decide adaptively the number of hidden layers and connection weights of a net, which are very important parameters for RBFNN. The method can improve the performance of SRR of satellite image and speed up the convergence of RBFNN to 221 s. Experimental results of simulation and generalization indicate that the well-trained RBFNN can realize the SRR of satellite images in higher spatial resolutions, higher efficiencies and lower errors.