Abstract: Image retargeting is one of the fundamental problems in image editing, which studies how to rearrange image content according to various application requirements. Current research on light field image retargeting except angular and spatial super-resolution is still blank. This paper proposes a method to edit the baseline of light field images. The method mainly contains three steps. First, the method calibrates the light field image to obtain the camera parameters, and estimates the disparity maps for each subaperture view. Then the method retargets the light field image according to the baseline editing requirement, which means that all subaperture views are projected to the corresponding subaperture viewpoint of target light field image. Finally, a deep learning method is proposed to optimize the target light field images, which can well inpaint the black holes of disoccluded regions in direct retargeting process. Experimental results demonstrate that the method can perform baseline editing based light field images retargeting, and obtain visual pleasure synthesized light field image. The proposed method can be used in a wide variety of light field images processing applications, such as light field image stitching, object copy and paste in different images, stereoscopic 3D display of light field images, etc.

Key words: light field image, baseline editing, image retargeting, deep convolutional neural network

摘要： 图像重定向是图像编辑中的一个基本问题，主要研究根据具体应用要求对图像内容进行重构。当前针对光场图像角度和空间超分辨率以外的重定向的研究尚属空白。提出了一种光场图像基线编辑方法。算法主要包含三个步骤：首先对光场图像进行标定得到相机参数，并估计光场图像每个子视点图像的视差图；然后根据基线编辑的要求对光场图像进行重定向处理，即将每个子视点图像投影到目标光场图像对应的子视点；最后构建一个深度学习算法对目标光场图像进行优化，对直接重定向过程中因遮挡去除导致的图像空洞区域进行修复。实验结果表明，所提算法能够实现基线编辑的光场图像重定向处理，得到高质量的目标光场图像。该算法可用于一系列光场图像编辑应用，包括光场图像拼接、不同图像间物体拷贝和复制、光场图像的立体显示等。

关键词: 光场图像, 基线编辑, 图像重定向, 深度卷积神经网络