Research On Depth Estimation Method Of Light-field Image

The light field camera can record multi-view information of a three-dimensional scene within one shot,which possesses unique advantage especially in depth estimation.However,existing depth estimation methods can not handle complex occlusion scene well,and the processing speed of them is slow,because the amount of light field image data is large.Therefore,this paper focuses on the research of light field depth estimation about the two difficult problems of occlusion and computational time complexity.The main work of this thesis is as follows:(1)We summarize the research status of 3D scene depth estimation,and expound the origin of light field theory,the acquisition method of light field information,the decoding process and visualization method of light field image,then analyze the light field characterization image obtained by three visualization methods and the basic principle of calculating the depth of the scene.(2)A method for estimating the depth of anti-occlusion light field based on Gini cost volume is proposed.Firstly,the refocusing images are obtained by the light field refocusing algorithm.Then,the Gini cost volume of the central view and other views are constructed.The initial depth map is calculated according to the principle of minimum cost volume.Finally,the initial depth information is combined with the color map for joint guided filtering,and a high-precision depth map is obtained.The experimental results show that the depth information obtained by this method is more accurate and the edge structure of the occlusion area is good.(3)We combine the convolutional neural networks to implement an end-to-end deep prediction network.Firstly,the multi-channel sub-network is used to extract the features of stack images in different directions.Then,the attention model is used to enhance the geometric relationship of the extracted features.Finally,the final scene depth information is obtained through a full convolution network.The experimental results show that the depth information obtained by the model has a small computational complexity while maintaining high precision.