传统虚拟视点生成采用像素填充法对生成的虚拟视点图像进行空洞填充和伪影修复, 其修复效果无法满足自由立体显示需求。为了获取高质量的虚拟视点图像, 提出了一种基于深度卷积神经网络的虚拟视点生成方法。该方法采用随机初始化的深度卷积神经网络作为图像先验, 经过卷积神经网络结构的不断迭代, 对虚拟视点图像的空洞和伪影进行修复, 并将得到的高质量虚拟视点图像合成为自由立体图像, 用于自由立体显示。修复后的虚拟视点图像的PSNR均值为25.6, 相比传统像素填充方法有明显提升。实验结果表明, 所提方法能够实现高质量的自由立体显示效果。

The conventional virtual viewpoint generation methods use the pixel filling algorithm to fill the image holes and restore artifacts on the generated virtual viewpoint images. However, the repaired virtual viewpoint images cannot meet the requirements of the autostereoscopic display. In order to obtain virtual viewpoint images with high quality, we propose a virtual viewpoint image generation method for autostereoscopic display based on deep convolutional neural network (DCNN). The method uses a randomly initialized deep convolutional neural network as an image prior, iterates through the structure of the convolutional neural network, and effectively repairs the image holes and artifacts of the virtual viewpoint image. The repaired virtual viewpoint images are then synthesized into the autostereoscopic image for autostereoscopic 3D display. The average value of PSNR of the virtual viewpoint images is 25.6, which is higher than that of the conventional pixel filling method. The experimental results demonstrate the proposed method can obtain high quality autostereoscopic 3D images.