Multi-view 3D Point Cloud Reconstruction Algorithm Based On DNN

3D modeling is the primary way people understand the world,while 3D reconstruction refers to the technique of restoring an object into a 3D model through 2D images from different angles.Imaged-based 3D reconstruction technology has significant application value and broad development prospects in fields such as medical imaging,game development,and military reconnaissance.The traditional 3D reconstruction method requires manual design of parameters and characteristic matrix,and the effect of 3D reconstruction varies greatly under different external environment.In recent times,a number of researchers have started to utilize deep learning technology in the area of 3D reconstruction,owing to its extensive use in computer vision.Early 3D reconstruction networks used in a relatively simple feature extraction method,which made it prone to problems such as lost detail features.When projecting the generated depth map into a 3D point cloud,the density of the point cloud was low,which resulted in holes and poor quality of the final 3D model.Considering the issues present in the previously mentioned 3D reconstruction algorithm,this thesis introduces the following enhancements:(1)This thesis uses a feature pyramid network in place of the conventional convolutional neural network for feature extraction,which has greater flexibility and can process images of varying scales more effectively.Additionally,a feature fusion module is introduced to regulate the transmission of information from deep to shallow layers,enhancing its capacity to extract intricate details.(2)To avoid the substantial memory consumption produced by traditional 3D convolutional networks in the depth map estimation module,this thesis proposes an adaptive depth sampling method for depth image estimation.This method generates cost volumes of different sizes to estimate the depth map and reduces the estimated range of the current depth map based on the depth map output from the upper layer.This approach reduces the resource consumption during the depth estimation phase.(3)For the point cloud processing module,this thesis proposes optimizing the selection of central points using a graph neural network based on the point cloud prediction with Point Flow.Firstly,the GRID-GCN algorithm is used to choose the appropriate target point,which allows the neighborhood around the target point to cover a larger spatial volume.Secondly,the Point Flow algorithm is used to predict the point cloud at the target point.Thirdly,the obtained point cloud is strengthened by obtaining point cloud patches that are controlled by the distance to the target point,combining the original point cloud patches with interpolated point cloud patches,and finally fusing them to obtain a complete and high-density 3D point cloud of the target scene.To validate the effectiveness of the above method,this thesis adopts the DTU datasets for model training and testing.Quantitative analysis is conducted on point cloud accuracy and completeness,algorithm resource consumption and runtime,and actual reconstruction quality,and traditional 3D reconstruction methods and existing deep learning-based 3D reconstruction algorithms are compared.The results show that compared to traditional 3D reconstruction methods,deep learning-based 3D reconstruction methods generally have better reconstruction effects,and compared to existing 3D reconstruction networks,the point cloud reconstructed by the algorithm proposed in this thesis has better processing effects on detail features,effectively improving the performance of 3D reconstruction algorithms.