Research On Light Field Image Reconstruction Algorithm Based On Multi-Representation Cooperation

Light field(LF)imaging has been widely used in low-level image processing,middlelevel visual analysis and high-level user interaction since scene spatial and angular information can be captured simultaneously.However,current LF cameras suffer from the inherent spatioangular tradeoff,and the limited angular resolution and spatial resolution hinder the development of LF imaging technology.To alleviate these problems,this thesis uses the multirepresentation collaboration of LF images to propose two reconstruction networks to perform angular and spatial reconstruction of LF images respectively.In the aspect of angle reconstruction of LF images,this thesis proposes the geometryassisted multi-representation view reconstruction network for LF images angular superresolution.The structural characteristics of multiple representations such as LF lenslet images,sub-aperture image arrays and pseudo video sequences are fully explored to reconstruct the LF images stack with high angular resolution.Then,in order to further improve the reconstruction performance,we construct a bidirectional view stack structure and a geometry-aware refinement network to fully integrate the intra-LF spatial-angular information and inter-LF geometry information.Compared with other angle reconstruction algorithms,the proposed method shows better reconstruction performance on both real-world and synthetic LF scenes.Furthermore,extensible applications on scene depth estimation also demonstrate that the proposed method can recover more texture details.In the aspect of spatial reconstruction of LF images,this thesis proposes the cross-arranged multi-stream feature fusion reconstruction network for LF images spatial super-resolution.The high-dimensional features of multi-representation forms of LF images are extracted and fused to construct a super-resolution module based on cross-arranged.Specifically,in this thesis,the high-dimensional features of the LF are fully mined in angle and space by combining subaperture images and epipolar plane images,and the extracted high-dimensional features are fused by the channel attention module.Then,in order to further explore the angle information,this thesis constructs a cross-arranged structure of angle views to further improve the quality of LF spatial reconstruction.Compared with other spatial reconstruction algorithms on five public datasets validate that,the proposed method achieves better reconstruction performance.