Research On Depth Estimation Algorithm Based On Light-Field Images

In recent years,with artificial intelligence technology rising,3D technologies such as Virtual Reality(VR),Augmented Reality(AR),stereoscopic 3D video,and 3D printing have developed rapidly,and various applications have emerged one after another.In many applications,how to obtain the depth information of the scene has always been the focus of research.The traditional camera can get the spatial information of the scene well,but the depth information is lost during the shooting process.The traditional method uses a single camera and a dual camera to estimate the depth of the scene.It has less views and the precision is low.The use of multiple cameras has more views and the precision is high,but it also has high complexity and high cost.So each method has some defects.With the advent of light field cameras,the technology of depth estimation has a new research hotspot.Compared with conventional cameras,a light field camera can achieve simultaneous information collection for both spatial and angular domains and also solve the contradiction between the number of views and the cost of the single,dual camera and multi-camera acquisition depth methods.Therefore,the use of light field cameras to estimate the depth of the scene has great advantages and research prospects.However,there are many challenges in real-world scenarios,such as the presence of occlusion phenomena,flat areas of texture and so on.These are some of the research difficulties in using the light field camera to estimate the depth of the scene.Based on this,this paper has carried out the research on these issues.Firstly,for the occlusion problem,this paper derives the relationship between the spatial information and the angular information in the light field from the illuminating path of the light field,and establishes the anti-occlusion model between the spatial information and the angular information.Based on this model,a corresponding views extraction algorithm is proposed for different occlusion situations,which realizes depth estimation under complex occlusion and improves the robustness of the algorithm.Then,for the noise problem of texture flat region,this paper proposes a fusion of confidence and global constrained Markov energy model to globally optimize the initial depth map.A curvature-based confidence calculation method is designed in the energy model to improve the optimization accuracy.Furthermore,considering the generalization ability of deep learning network and the long calculation time of traditional algorithms,this paper also proposes a multi-channel light field depth estimation algorithm based on neural network.Based on the criteria about different slopes on the epipolar plane corresponding to different depth,the algorithm designs the input images under different slopes and extracts the angular features of the light field image into the multi-channel network.Then the high-dimensional features are extracted through the deep convolutional neural network,and the completed depth map is finally obtained.The algorithm can improve the accuracy of depth estimation and reduce time overhead.