Shape Constraint-based Multi-view Object Reconstruction Algorithm And Parallel Research And System Implementation

Reconstructing object from multiple views as 3D point cloud is a challenging task in the field of computer vision and 3D reconstruction.Complete,accurate and small model reconstruction of object from multiple views is essential for applications in vision fields such as autonomous driving and artificial intelligence.It is still quite difficult to reconstruct the corresponding object completely and accurately from multiple views,and when trying to obtain the feature of object from a series of object images,the object feature obtained can be inaccurate or even fail to obtain the feature information corresponding to a particular pixel due to the changes in luminosity and background resulting from multiple views taken from different angles of the same camera.Most of the current multi-view reconstruction methods usually extract multiple views including a reference view and multiple source view features into a single-image based feature extraction network to obtain a set of feature information,then all the feature information is transformed by differentiable homography to construct a cost volume based on the reference view perspective,and finally use the U-Net regression model to output the depth map of the object and use the depth fusion module to obtain a 3D point cloud of the object.Existing methods improve on the classical multi-view network to some extent,but the 3D point cloud reconstructed from the object still suffers from missing edges and sparse structure.In this paper,we propose a new deep learning-based multi-view reconstruction method to address the problem of missing edges and sparse structure caused by photometric inconsistencies under multi-view shooting conditions.Although these works are implemented based on the classical MVSNet network and have achieved good results with improvements,it is difficult to find a balance between accuracy and model size and parameters.In addition,the large datasets and complex network modules in multi-view 3D reconstruction tasks can make the training period long.In this paper,we design a training method based on a parallel acceleration algorithm,which can effectively reduce the training time to improve efficiency.The main work of the paper is as follows:1.To allow the model to acquire more detailed edge structure feature utilizing a self-supervised approach,this paper proposes a multi-view reconstruction method based on an edge structure extractor to fill in the multi-view features using edge structure feature to make the final feature information better and effectively reconstruct the 3D point cloud of the object.In this paper,the edge structure map of the source view is transformed to the viewpoint of the reference view to provide self-supervised information for the edge structure map of the reference view.In addition,a combined local and global masking mechanism is designed in this paper to ensure that a richer masking enhancement consistency is provided to enable the model to achieve some accuracy even on views with low lighting.2.In order to improve the training speed of multi-view object 3D reconstruction,this paper designs a parallel-based network for model training acceleration of multi-view reconstruction task.Distributed data parallelism is used to parallelise multiple processes on multiple graphics card nodes,solving the load balancing problem between graphics card and improving the utilisation of computing resources.Moreover,the sub-network modules of the network operate in parallel on different graphics cards using a pipeline model to alleviate the problem of large model memory occupation for 3D reconstruction of multi-view object.3.Build and implement a multi-view 3D reconstruction system with Bootstrap framework for GUI interface,so that the algorithm proposed in this paper can be applied in a systematic way on the Web platform,which can realize 3D reconstruction of multi-view object and visualize them.