Research And System Implementation Of Multi-view Stereo Based On Dynamic Edge Flow And Tensor Acceleration

In the past few years,the rise of large-scale deep learning and the urgent need for the implementation in the industry has led to a proliferation of related technologies in computer vision,as one of them,the computer 3D reconstruction technology has also developed rapidly.As an important carrier of spatial information,three-dimensional models are able to convey more intuitive and rich geometric information than two-dimensional graphics.And computer 3D reconstruction technology is a very challenging task in the process of using two-dimensional graphics to restore three-dimensional models.Traditional 3D reconstruction technology started early with in-depth research,mature technology and extensive applications.It has received much attention in the areas such as medical equipment,cultural relic models and virtual reality.While the emerging 3D reconstruction technology is based on deep learning models,which uses the powerful predictive capabilities of deep learning to process some of the key steps,and generally achieves better reconstruction results than the traditional methods.Most of the existing deep network-based multi-view 3D reconstruction methods are dedicated to improving the accuracy of the reconstructed model.However,due to the external factors such as complex scenes,occlusion,and lighting,the reconstructed 3D model is too sparse especially in areas with ambiguity in depth such as the edge of the scene.In addition,deep network-based multi-view 3D reconstruction methods generally require high computational cost and memory consumption,which makes the real-time 3D reconstruction and portable device transplantation more difficult.To deal with these problems,this paper proposes a multi-view depth inference method based on dynamic edge flow which greatly increases the density of the reconstructed model.At the same time,tensor decomposition is used to accelerate the study of three-dimensional convolution which reduces the algorithm running time and memory consumption to a large extent.This dissertation focuses on the following researches:1.Considering that traditional 3D reconstruction methods often have low prediction accuracy at the edge of the scene which directly leads to sparse reconstruction,a high-resolution depth inference method based on dynamic edge flow is proposed.This method dynamically flows for complex areas such as the edge of the scene in the depth map,so as to obtain better edge details and continuously impro ve the resolution of the reconstructed depth map and finally realize the reconstruction of a dense threedimensional model.2.The cost-volume-based 3D reconstruction methods use a number of three-dimensional convolution(3DCNN)regularization costs,which leads to a slow reconstruction process and extremely memory consumption.In response to this problem,this paper uses tensor decomposition(CP decomposition)to accelerate the study of three-dimensional convolution,and uses the mathematical principle of tensor decomposition to solve three-dimensional convolutions with 5D dimensions into several rank-one convolutions to achieve network dimensionality reduction.Experimental results prove that the tensor decomposition can effectively improve computational efficiency and reduce memory consumption.3.According to the above research,a web page system is designed for real-time multi-view 3D reconstruction and display of 3D models.