一种基于深度网络的视图重建方法

A View Reconstruction Method Based on Deep Network

为了解决在仅有单目视图的环境下实现立体匹配的问题,在现有视图重构网络模型Deep3D的基础上,提出了基于加权局部对比归一化约束的全卷积重构模型.该模型采用改进的全卷积神经网络架构作为模型的特征提取模块,以期减少训练参数,降低训练时间,增加模型的非线性.为了进一步提高重构精度,设计了新的基于加权局部对比归一化的约束条件,并采用结构相似性成本(SSIM)与L 1成本相结合的损失优化函数对模型进行优化.在KITTI 2015数据集上展开实验,并与Deep3D模型及其后续的改进方法进行比较.实验结果表明,在只使用左视图作为训练数据的情况下,生成的右视图在SSIM和峰值信噪比两个指标上有很大提升,能够满足立体匹配方法中右视图的精度要求.

To deal with stereo matching in the environment of only a single view,a full convolution reconstruction model with weighted local contrast normalization constraint is proposed on the basis of the existing view reconstruction network model Deep3D.This model adopts the improved full convolutional neural network architecture as the feature extraction module of the model to reduce the training parameters and training time,and to increase the nonlinearity of the model.In order to further improve the accuracy of reconstruction,a new constraint condition based on weighted local comparison normalization is designed,and a loss optimization function combining structural similarity(SSIM)cost and L1 cost is used to optimize the model.Experiments were carried out on the KITTI 2015 dataset,and compared with the Deep3D model and subsequent improvements.The experimental results show that the generated right view has a great improvement in SSIM and peak signal to noise ratio when only the left view is used as the training data,which can meet the accuracy requirements of the right view in the stereo matching method.