Research On 3D Image Reconstruction Algorithm Based On Multi-View

Multi-view stereo(MVS)is one of the most challenging tasks in computer vision,which restores three-dimensional information of objects from a set of images with known camera parameters.After years of development,3D reconstruction has been widely applied in medicine,the military,industry,and other fields.At present,the main research direction of 3D image reconstruction based on multi-view is towards lightweight development.In early research,the cost volume was constructed through 3D convolution which required a significant amount of computational resources and time although structurally simple.Therefore,how to reduce the consumption of computing resources without reducing the reconstruction accuracy has become the focus of research;Simultaneously,the model’s generalization ability is not good.How to improve the generalization ability and the integrity of the model is also a research difficulty.Given the above problems,the main research contents are as follows:(1)In order to reduce computational consumption,this thesis proposes a multi-view3 D reconstruction algorithm based on dynamic iterative cost volume.This algorithm utilizes the construction of dynamic cost volumes to reduce computational costs caused by complex 3D convolutions.By fusing the geometric features,depth features,and neighborhood frame image mapping features of the reference frame image,a dynamic cost volume is constructed.Compared with other works,the dynamic cost volume is smaller.Therefore,it can be input into the Gate Recurrent Unit(GRU)for multiple iterations to optimize the depth map and improve the reconstruction accuracy;Layer structure and multi-scale feature extraction speed up model convergence.The experimental results show that the dynamic cost volume method can shorten the calculation time and reduce memory consumption without affecting accuracy.(2)In order to improve the generalization ability of the model,this thesis proposes a multi-view patchmatch algorithm based on data augmentation.In the data input stage,data augmentation is utilized to change the brightness,contrast,and hue of the input image.Simultaneously removing random area pixels from the image and simulating the presence of highlights and occlusion in the image,lays the foundation for improving model robustness during subsequent network model training.In the depth estimation stage,the dynamic sampling interval method is used to obtain more image information.Simultaneously utilizing the feature information of neighboring pixels,a cost volume is constructed by a patchmatch algorithm to achieve the optimal estimation of depth values.Through adaptive propagation and adaptive evaluation modules instead of traditional 3D convolution,it can effectively improve the accuracy of reconstruction while reducing memory consumption.Through comparative experiments,it is verified that the 3D point cloud reconstructed by this algorithm has higher integrity in the DTU dataset and the generalization performance verification on the Tanks & Temples dataset shows that this algorithm has strong generalization ability.