Research On Light Field Depth Estimation For Complex Scenes

Light field images(LFI)can not only record rich light information such as the position and angle of light,but also provide more immersive perception by increasing spatial resolution and the number of views.It has attracted more and more attention in recent years.Depth estimation is the research focus in the field of computer vision.Accurate depth estimation is of great significance to production and life.Light field depth estimation can be applied to many fields such as industrial model building,3D printing,micro-scanning,remote sensing mapping,face recognition,VR image and so on.It has broad development prospects in the information age,so the research on light field depth estimation is very necessary.Because the light field image contains rich angle information,it has a natural advantage for depth estimation,but there are often occluders in complex scenes,which makes it easy to make the depth estimation of the occlusion edge area inaccurate.In view of this problem,from the perspective of coping with occlusion,this thesis conducted an in-depth study of the light field depth estimation of complex scenes.The main work and results of this thesis are as follows:1.Research on the problem of occlusion edge estimation in complex scenes.After analyzing and comparing Canny edge detection and LK optical flow method,this thesis proposes a fusion edge detection method.Based on the complementary characteristics of Canny edge detection and LK optical flow method edge detection,the two are fused.The result of the LK optical flow method is retained in the texture area of the object and replaced by the Canny result at the edges.The experimental results show that this fusion method avoids the false detection of the internal texture area as much as possible,and improves the reliability of edge estimation.2.Research on the occlusion edge angular patches and spatial patches partitioning in complex scenes.This thesis analyzes and compares K-means clustering and hierarchical clustering methods.An improved hierarchical agglomerative clustering method is proposed.Outlier detection is added to the HAC,and a similarity measure based on the average degree of attraction between clusters is constructed.The final clustering result is determined by optimizing the average degree of attraction between clusters and the degree of intra-cluster concentration.Improved HAC solves the problem that K value and random cluster center selection make the clustering result unstable and the need to select the number of hierarchical clusters according to the occlusion situation.The impact of outliers is also reduced.3.The fusion edge detection and improved hierarchical agglomeration clustering method proposed in this thesis is applied to the light field image depth estimation in complex scenes,and the accurate depth map is obtained by MRF optimization.The final experimental results show that compared with several current popular methods,the depth estimation mean square error of the method in this paper is the smallest,which is an average drop of 32% compared with the advanced method,which is more accurate.Especially the occlusion edge area and the internal texture area,the number of false detections is significantly reduced,which illustrates the effectiveness of the proposed method.In order to verify the significance of the module in this article,we conduct an ablation experiment.The results show that each module has a positive effect on the depth estimation results.