Abstract:

To solve the problem of low accuracy and time-consuming computation in traditional depth estimation methods of light field, a densely connected CNN which has multiple input stream to estimate depth of light field is proposed. First, this algorithm preprocesses input images and converts them into Epipolar Plane Image volume（EPI）. Secondly, it uses light field specific data augmentation methods like random gray scale to overcome the lack of training data. Then, the CNN can convert all features into depth map. This algorithm uses shortcuts structures to reduce computation. Experimental results on HCI 4D light field dataset show that densely connected CNN can achieve better results in mean squared error and bad pixel ratio, and the efficiency is improved a lot.

Key words: light field, depth estimation, Epipolar Plane Image（EPI）, Convolutional Neural Network（CNN）, data augmentation, shortcuts

摘要：

针对传统的光场深度估计算法精度低、计算慢的问题，提出了一种改进DenseNet的多输入流密集连接型卷积神经网络进行光场深度估计的方法。该方法采用的密集连接的结构，减少了模型的计算量。对输入图片进行预处理，转化为极平面图EPI Volume（Epipolar Plane Image）结构，采用随机灰度化等数据增强方法克服训练数据不足，通过神经网络将EPI特征转化为深度信息。在HCI 4D光场数据集上的对比实验结果表明，该方法在均方误差和不良像素率上都取得了良好结果，并且在执行时间上大幅领先于传统算法。

关键词: 光场, 深度估计, 极平面图, 卷积神经网络, 数据增强, 密集连接型网络