【作者】 王奇； 傅雨田；

【Author】 Wang Qi;Fu Yutian;Key Laboratory of Infrared System Detection and Imaging ,Chinese Academy of Sciences;Shanghai Institute of Technical Physics,Chinese Academy of Sciences;University of Chinese Academy of Sciences;

【机构】 中国科学院红外探测与成像技术重点实验室； 中国科学院上海技术物理研究所； 中国科学院大学；

【摘要】 本文提出一种单幅图像动态重聚焦的方法,结合基于深度学习的光场合成与基于几何结构的弥散圆渲染方法,模拟光场重聚焦效果,渲染不同的浅景深视图。以往的计算成像方法直接使用滤波函数对单幅图像的前景和背景离焦渲染或者使用多视角子图提供立体信息分层渲染。本文仅输入单幅图像,首先进行深度估计,然后将深度图转换为视差信息,在不同深度确定弥散圆尺寸对像素重采样。设计了两种神经网络结构,分别以光场相机的多视角子图和重聚焦图像为样本进行有监督的深度学习。我们在多个数据集进行验证,包含了实际拍摄的场景。实验表明,我们的方法在可接受的计算成本获得了优于其他方法的视觉效果与评价指标,PSNR达到了34.17,SSIM达到了0.933。更多还原

【Abstract】 We present a method to dynamically refocuse a single image,combining the light field synthesis based on deep learning with the circle of confusion rendering method based on geometric structure,simulating the light field refocusing effect and rendering different depth of fields(DOFs).Prior computational imaging methods directly use the filter function to render the foreground and background of a single image or use multiple views with stereo information for layered rendering.We instead use a single image,depth estimation is performed firstly,the depth map is converted into disparity,finally the CoC diameter is measured at different depths to resample pixels.Two neural network structures are designed,supervised by multi-views and the refocused DOFs of the light field.We conduct experiments on public datasets and scenes shot by mobile phone.Results show that our method achieves superior visual performance and evaluation indicators than other methods with acceptable computational cost,PSNR reaching 34.17 and SSIM reaching 0.933.更多还原