| In recent years,methods based on deep learning have shown strong competitiveness in the multi-view 3D reconstruction depth estimation.According to the principle of camera imaging,any point on the perspective line corresponds to the same imaging point on the imaging plane,and its corresponding point in 3D space cannot be uniquely determined according to the pixel points on the single image.The multi-view 3D reconstruction algorithm could eliminate this kind of uncertainty.Different views in multi-view data are different descriptions of the same target object,which could also complement information,and each view has different concerns.Besides,multi-view could get a more comprehensive feature representation of the target object.In the existing multi-view 3D reconstruction algorithm,the methods based on constructing 3D cost volume capture the geometric structure of the object and reduce the impact of image distortion caused by perspective transformation and occlusion.However,the boundary of the reconstructed depth map may be overly smooth due to the large sensitivity field involved in the three-dimensional convolutional neural network.The 3D convolutional neural network is time-consuming in video memory and operation time,which limits the output of high-resolution depth map to some extent.Human visual system has a kind of ability which choose the most important part of observed information from the complex visual scene,ignore the other minor information.Inspired by this ability and in order to make full use of limited computational resources and huge calculated quantity of deep neural network,this paper proposes the Attention-guided multi-view network for depth estimation.The algorithm uses MVSNet to predict a low-resolution rough depth map.At the same time,considering that high-quality texture features which are important for multi-view 3D reconstruction,the reference images in the natural environment contain the edge details needed in the reconstruction process.Therefore,this paper used the U-Net framework to extract the detailed feature information of different scales of the reference image to guide the refinement process of the depth map.Finally,the 3D point cloud model corresponding to the image is obtained by depth map fusion and denoising.The main work of this paper is as follows:1.Combining camera geometry and deep learning,a convolution neural network is used to capture 2D image features of different shooting angles.Using the extracted image features and the corresponding camera parameters,we could generate the rough depth map of 3D structure of objects through the dense matching and 3D-CNN regression.Since the reference image in the natural environment contains the edge details of the reconstructed target,this paper combines the texture details information from the reference image and the rough depth map containing 3D geometric information,use the attention network to guide the refinement of the rough depth.2.Based on the view-based multi-depth map fusion algorithm,the proposed method integrates depth maps from different angles into 3D space to generate 3D point cloud model.In order to improve the accuracy and completeness of the reconstruction of the point cloud model,we use probability maps to measure the confidence in the prediction of the depth map filtering out the noise points whose confidence in the predicted depth map is below the threshold.The depth map is fused to form the target point cloud model through the constraint of depth consistency criterion among depth maps from different perspectives.3.Due to a large amount of data and the 3D convolutional neural network,the network training cost is huge.Through the research and analysis of the parallel algorithm of deep learning model,the time-consuming part of the algorithm is optimized with CUDA technology based on GPU-acceleration to optimize the multi-view 3D reconstruction algorithm,which improved the convergence speed of neural network,and reducing the time cost of network training.The experimental results show that the reconstructed surface of the algorithm in this paper is more complete,especially generating more detailed point clouds in the areas of low texture,reflected surface,and repeated texture of the image,which has better reconstruction effect.The time-consuming part of the algorithm that can be executed in parallel is optimized by GPU-based CUDA parallel programming technology,which further improves the execution efficiency of the 3D reconstruction algorithm and obtains training results quickly. |
