利用ANNF统计预测模型的单帧超分辨率算法  

A Single Frame Super-resolution Algorithm Using ANNF Statistical Forecasting Model

作　　者：徐亮[1] 李欣[1] 

机构地区：[1]新疆工程学院计算机工程系,乌鲁木齐830000

出　　处：《控制工程》2017年第9期1918-1924,共7页Control Engineering of China

基　　金：新疆维吾尔自治区高校科研计划青年教师科研启动基金(XJEDU2016S085);2016年度新疆工程学院科研基金(无人机航拍图像超分辨率重建算法研究)

摘　　要：针对基于稀疏不变性假设的单帧超分辨率(SR)算法的局限性,提出一种利用相似最近邻(ANN)统计预测模型的单帧SR算法。首先,利用相似最近邻思想,通过波尔茨曼机捕捉HR字典与LR字典对稀疏模式之间的依赖关系,建立统计预测模型;然后,根据LR块与HR块相关的最小均方误差(MMSE)计算网络参数,获得它们的依赖关系;最后,利用多层前向神经网络提取字典元素内积,通过计算重叠局部块预测值的均值来重建图像。利用峰值信噪比PSNR和结构相似性度量SSIM评估实验结果,实验结果表明,提出的算法在视觉效果和数值标准方面大多优于其他算法,在选择合适参数情况下,峰值信噪比至少提高0.2 d B。For the limitations of single frame super-resolution algorithm based on sparse invariance, a single frame super-resolution algorithm using a statistical forecasting model of approximate nearest neighbor field is proposed. Firstly, dependencies of sparse pattern between HR and LR dictionary can be captured by Boltzmann machine and using the thought of approximate nearest neighbor field, the statistical forecasting model is built. Then, the network parameters are calculated by the minimum mean square error （MMSE） of the LR pitches and HR pitches, and their dependencies are accessed. Finally, multilayer neural network is used to extract the products of elements in the dictionary. The image is reconstructed by means of the predicted values with the calculation of local overlap. Peak signal to noise ratio （PSNR） and structural similarity measure （SSIM） are used to assess the results, the experimental results show that the proposed algorithm is mostly better than that of others in visual and numerical standards. PSNR can be up by at least 0.2 dB in the case of selecting the appropriate parameters.

关 键 词：超分辨率重建 稀疏不变性 相似最近邻 统计预测模型 神经网络 最小均值误差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]