The point spread function(PSF) is investigated in order to study the centroids algorithm in a reverse Hartmann test(RHT) system. Instead of the diffractive Airy disk in previous researches, the intensity of PSF be...The point spread function(PSF) is investigated in order to study the centroids algorithm in a reverse Hartmann test(RHT) system. Instead of the diffractive Airy disk in previous researches, the intensity of PSF behaves as a circle of confusion(CoC) and is evaluated in terms of the Lommel function in this paper. The fitting of a single spot with the Gaussian profile to identify its centroid forms the basis of the proposed centroid algorithm. In the implementation process, gray compensation is performed to obtain an intensity distribution in the form of a two-dimensional(2D) Gauss function while the center of the peak is derived as a centroid value. The segmental fringe is also fitted row by row with the one-dimensional(1D) Gauss function and reconstituted by averaged parameter values. The condition used for the proposed method is determined by the strength of linear dependence evaluated by Pearson's correlation coefficient between profiles of Airy disk and CoC. The accuracies of CoC fitting and centroid computation are theoretically and experimentally demonstrated by simulation and RHTs. The simulation results show that when the correlation coefficient value is more than 0.9999, the proposed centroid algorithm reduces the root-mean-square error(RMSE) by nearly one order of magnitude, thus achieving an accuracy of - 0.01 pixel or better performance in experiment. In addition, the 2D and 1D Gaussian fittings for the segmental fringe achieve almost the same centroid results, which further confirm the feasibility and advantage of the theory and method.展开更多
目的传统的基于子视点叠加的重聚焦算法混叠现象严重,基于光场图像重构的重聚焦方法计算量太大,性能提升困难。为此,本文借助深度神经网络设计和实现了一种基于条件生成对抗网络的新颖高效的端到端光场图像重聚焦算法。方法首先以光场...目的传统的基于子视点叠加的重聚焦算法混叠现象严重,基于光场图像重构的重聚焦方法计算量太大,性能提升困难。为此,本文借助深度神经网络设计和实现了一种基于条件生成对抗网络的新颖高效的端到端光场图像重聚焦算法。方法首先以光场图像为输入计算视差图,并从视差图中计算出所需的弥散圆(circle of confusion,COC)图像,然后根据COC图像对光场中心子视点图像进行散焦渲染,最终生成对焦平面和景深与COC图像相对应的重聚焦图像。结果所提算法在提出的仿真数据集和真实数据集上与相关算法进行评价比较,证明了所提算法能够生成高质量的重聚焦图像。使用峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性(structural similarity,SSIM)进行定量分析的结果显示,本文算法比传统重聚焦算法平均PSNR提升了1.82 d B,平均SSIM提升了0.02,比同样使用COC图像并借助各向异性滤波的算法平均PSNR提升了7.92 d B,平均SSIM提升了0.08。结论本文算法能够依据图像重聚焦和景深控制要求,生成输入光场图像的视差图,进而生成对应的COC图像。所提条件生成对抗神经网络模型能够依据得到的不同COC图像对输入的中心子视点进行散焦渲染,得到与之对应的重聚焦图像,与之前的算法相比,本文算法解决了混叠问题,优化了散焦效果,并显著降低了计算成本。展开更多
基金Project supported by the National Natural Science Foundation of China(Grant No.61475018)
文摘The point spread function(PSF) is investigated in order to study the centroids algorithm in a reverse Hartmann test(RHT) system. Instead of the diffractive Airy disk in previous researches, the intensity of PSF behaves as a circle of confusion(CoC) and is evaluated in terms of the Lommel function in this paper. The fitting of a single spot with the Gaussian profile to identify its centroid forms the basis of the proposed centroid algorithm. In the implementation process, gray compensation is performed to obtain an intensity distribution in the form of a two-dimensional(2D) Gauss function while the center of the peak is derived as a centroid value. The segmental fringe is also fitted row by row with the one-dimensional(1D) Gauss function and reconstituted by averaged parameter values. The condition used for the proposed method is determined by the strength of linear dependence evaluated by Pearson's correlation coefficient between profiles of Airy disk and CoC. The accuracies of CoC fitting and centroid computation are theoretically and experimentally demonstrated by simulation and RHTs. The simulation results show that when the correlation coefficient value is more than 0.9999, the proposed centroid algorithm reduces the root-mean-square error(RMSE) by nearly one order of magnitude, thus achieving an accuracy of - 0.01 pixel or better performance in experiment. In addition, the 2D and 1D Gaussian fittings for the segmental fringe achieve almost the same centroid results, which further confirm the feasibility and advantage of the theory and method.
文摘目的传统的基于子视点叠加的重聚焦算法混叠现象严重,基于光场图像重构的重聚焦方法计算量太大,性能提升困难。为此,本文借助深度神经网络设计和实现了一种基于条件生成对抗网络的新颖高效的端到端光场图像重聚焦算法。方法首先以光场图像为输入计算视差图,并从视差图中计算出所需的弥散圆(circle of confusion,COC)图像,然后根据COC图像对光场中心子视点图像进行散焦渲染,最终生成对焦平面和景深与COC图像相对应的重聚焦图像。结果所提算法在提出的仿真数据集和真实数据集上与相关算法进行评价比较,证明了所提算法能够生成高质量的重聚焦图像。使用峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性(structural similarity,SSIM)进行定量分析的结果显示,本文算法比传统重聚焦算法平均PSNR提升了1.82 d B,平均SSIM提升了0.02,比同样使用COC图像并借助各向异性滤波的算法平均PSNR提升了7.92 d B,平均SSIM提升了0.08。结论本文算法能够依据图像重聚焦和景深控制要求,生成输入光场图像的视差图,进而生成对应的COC图像。所提条件生成对抗神经网络模型能够依据得到的不同COC图像对输入的中心子视点进行散焦渲染,得到与之对应的重聚焦图像,与之前的算法相比,本文算法解决了混叠问题,优化了散焦效果,并显著降低了计算成本。