基于残差神经网络的高强度运动超分辨率图像重构

针对实际拍摄的亚像素信息较少的低分辨率运动图像,重构图像通常较为模糊,甚至不能分辨。为此,提出一种新的基于残差神经网络的高强度运动超分辨率图像重构方法。令沿运动方向的亮度保持恒定,通过光流场匹配实现高强度运动图像的运动估计;根据运动估计结果和超分辨率重构的基本思想,将BP神经网络看作残差神经网络的基础建立残差神经网络,对残差神经网络进行训练,参照训练样本将经插值法放大若干倍的待重构高强度运动图像作为输入,将高分辨率图像和输入图像间的残差作为输出,把输入和输出累加获取超分辨率图像,实现若干放大倍数高强度运动超分辨率图像的重构。实验结果表明,所提方法运动估计准确,重构图像清晰、质量佳。

Super-resolution image reconstruction of high intensity motion based on residual neural network

The reconstructed image is usually blurred and even can not be resolved in view of the low resolution moving image with less sub-pixel information. To this end, a new method of high - intensity motion super-resolution image reconstruction based on residual neural network is proposed. The brightness along the direction of motion is kept constant by optical flow field matching high intensity motion image motion estimation; motion estimation results and according to the basic idea of super-resolution reconstruction based on BP neural network, the neural network is established as the residual error of neural network, the training of residual neural network training samples, reference by interpolation method amplified several times to reconstruct high intensity motion image as input, the residual high resolution image and the input image as the output, input and output to the cumulative acquisition of super resolution image, to achieve a number of high magnification image super resolution reconstruction of motion intensity. The experimental results show that the proposed method is accurate in motion estimation, and the reconstructed image is clear and good in quality.