Abstract:

To solve the problem of low accuracy in traditional depth estimation methods of light field which is caused by using only one depth clue, a multi-input convolutional neural network combining disparity and defocus is proposed. This algorithm builds two depth clues by converting input images into EPI Volume and Refocus Volume, using the CNN to combine the EPI and defocus features into depth map. To overcome the lack of training data, data augmentation methods such as color transform, zoom, random gray scale and so on are adopted. Tests on HCI 4D light field benchmark dataset show that this method can achieve better result than traditional methods do in Mean Squared Error（MSE） and bad pixel ratio, also it is faster than traditional methods.

Key words: light field, depth estimation, Epipolar Plane Image（EPI）, disparity, defocus, Convolutional Neural Network（CNN）

摘要：

针对传统的光场深度估计算法采用单一的深度线索进行深度估计，导致估计结果精度较低的问题，，提出了一种融合视差和散焦量两种深度线索的光场深度估计的方法。该方法首先进行深度线索的构建，将输入的光场图片转化为EPI Volume和Refocus Volume结构，通过卷积神经网络将EPI特征和散焦特征转化为深度信息，为了克服训练数据不足，采用色彩变换、缩放和随机灰度化等方式进行数据扩容。最终在HCI 4D光场标准数据集上的对比测试结果表明，该方法在均方误差和坏像素率上优于传统算法，且在执行时间上也具有很大优势。

关键词: 光场, 深度估计, 极平面图, 视差, 散焦, 卷积神经网络