基于卷积神经网络的视差图生成技术

Disparity map generation technology based on convolutional neural network

针对裸眼三维中视差图生成过程中存在的高成本、长耗时以及容易出现背景空洞的问题,提出了一种基于卷积神经网络(CNN)学习预测的算法。首先通过对数据集的训练学习,掌握数据集中的变化规律;然后对输入卷积神经网络中的左视图进行特征提取和预测,得到深度值连续的深度图像;其次将预测所得到的每一个深度图和原图进行卷积,将生成的多个立体图像对进行叠加,最终形成右视图。仿真结果表明:该算法的像素重构尺寸误差相比基于水平视差的三维显示算法和深度图像视点绘制的算法降低了12.82%和10.52%,且背景空洞、背景粘连等问题都得到了明显改善。实验结果表明,卷积神经网络能提高视差图生成的图像质量。

Focusing on the issues such as high cost, long time consumption and background holes in the disparity map in naked-eye 3D applications, learning and prediction algorithm based on Convolutional Neural Network （CNN） was introduced. Firstly, the change rules of a dataset could be mastered through training and learning the dataset. Secondly, the depth map with continuous lasting depth value was attained by extracting and predicting the features of the left view in the input CNN. Finally, the right view was produced by the superposition of diverse stereo pairs after folding the predicted depth and original maps. The simulation results show that the pixel-wise reconstruction error of the proposed algorithm is 12.82% and 10.52% lower than that of 3D horizontal disparity algorithm and depth image-based rendering algorithm. In addition, the problems of background hole and background adhesion have been greatly improved. The experimental results show that CNN can improve the image quality of disparity maps.