Research On 3D Point Cloud Reconstruction Algorithm Based On Depth Image Prediction

Three-dimensional reconstruction refers to the extraction and analysis of the features of a photograph of a target object to restore a three-dimensional model of the object.Traditional 3D reconstruction methods are difficult to achieve a balance between performance and effect,and are seriously affected by external factors such as ambient lighting and sampling angle.Moreover,the feature matrix designed by hand is difficult and has poor robustness,while the convolutional network can effectively extract and integrate the feature information of the image.The early 3D reconstruction algorithm based on deep learning directly predicted 3D models in the form of point cloud and voxel through deep network,but the effect was poor.The MVSNet network proposed a 3D reconstruction solution by first estimating depth image and then creating point cloud with traditional algorithms.The network can extract and match the features of the low-dimensional images and avoid the problem that the features of the high-dimensional models are difficult to learn.Most of the subsequent reconstruction algorithms adopt the form of depth map prediction for model reconstruction,but the optimization scheme focuses on the reduction of network computation and ignores the effect of feature extraction on the reconstruction results.Therefore,based on the disadvantages of the three dimensional reconstruction algorithm using deep learning,this paper completed the following content: the depth of the existing network reconstruction of feature extraction module adopts convolutional network structure is relatively simple,only the space feature extraction and integration of the channel characteristics between using only simple addition to consolidate,ignoring the characteristics between the channel's impact on the final result.Attention mechanism has a good applicability for 3D reconstruction,adding attention mechanism can improve the attention of the target object of the network.In the reconstruction task,if the feature extraction module cannot effectively separate the features of the target object and the external background,it is difficult to obtain accurate reconstruction results.Based on the existing feature extraction module,this paper compares two attention mechanisms,SENet and CBAM,and verifies the effectiveness of the attention mechanism through experiments.In addition to the feature extraction module,the existing network depth map optimization module also has a large room for improvement.The MVSNet depth chart optimization module has shallow layers and small functions.RMVSNet uses cyclic network to generate and optimize the depth map,and the accuracy of the reconstruction results is improved,but the integrity is damaged greatly.The Point MVSNet selects the network structure consistent with the feature extraction module to optimize the depth map and has achieved certain improvement.Based on the depth map optimization module of Point MVSNet,this paper uses the residual network to replace the original general convolution structure,deepens the number of layers of the network,extracts more accurate local information,avoids the gradient disappearance of deep network,improves the effect of depth map optimization,and thus improves the quality of the final reconstruction results.In order to verify the effectiveness of the improved scheme,the commonly used3 D reconstruction data set was used for network training and experiment,and the influence of two fusion parameters,depth reliability and geometric error,on the reconstruction effect of the optimized network was tested.The feasibility of the proposed algorithm is proved by evaluating the reconstructed point cloud model results.