3D Scene Restoration Based On Weak Random Camera Pose Image

3D scene restoration involves extracting surface information from real scenes to create 3D models of target scenes.The multi-view-based scene restoration method is popular in various industries,such as auxiliary manufacturing,medical diagnosis,and education,due to its affordability and user-friendliness.The process consists of four main parts: acquiring twodimensional images,extracting and matching image features,calculating camera poses,and restoring sparse 3D point clouds,and dense 3D point clouds.Many scholars focus on improving3 D scene restoration by optimizing deep learning networks,and the camera pose distribution of the training dataset used in the experiments has the inherent characteristic of high normality.However,in practical applications,when ordinary users shoot the target scene,the camera pose distribution has a large randomness,which makes it difficult to ensure the acquisition of scene image data with equal or close quality to the training dataset,thus affecting the restoration effect.In order to alleviate the above problems,the following improvements are made in this paper.First,a weakly random camera pose image dataset is produced.This paper is oriented to the practical application of restoring 3D scenes,exploring the multi-view acquisition method in the field of optimized views photogrammetry,combining the characteristics of camera poses distribution of public data datasets and self-harvested datasets,and finally proposing a series of suggestions for acquiring multi-view images of target scenes.By providing users with suggestions for target scene capture,the randomness of camera pose distribution of acquired images is reduced,thus improving the effect of 3D restoration of target scenes.Second,the 3D scene restoration network is improved and optimized.The normalized camera poses in the public dataset are appropriately reduced so that they have weak normality and good 3D restoration effect at the same time.The depth restoration network AA-RMVSNet is used as the basic network framework of this paper,and the lightweight dual-channel attention module CBAM is integrated,which can improve the detailed performance of image features in channel and space,and realize the refined feature fusion,so as to effectively improve the completeness and accuracy of 3D point clouds.In conclusion,by using the target scene capture proposal proposed in this paper,the success rate of 3D scene restoration for the user-captured dataset is improved from 28.93% to81%.Using the randomized dataset for training and validation of this paper’s network,the experimental results show that the average accuracy error of the final recovered scene 3D point cloud is 0.377 mm and the average completeness error is 0.331 mm.Compared with AARMVSNet,the completeness is improved by 0.13% and the comprehensive index is improved by 0.09%.And it also has It also has more obvious advantages in terms of running time and memory consumption.Finally,the training effect of this paper’s network is tested using the multi-view images acquired in this paper.Compared with the traditional method COLMAP and the deep learning method MVSNet,this paper’s method has good performance in both the 3D visual effect and the experimental result data.In summary,the method in this paper provides a reference for ordinary users to capture technology and improve on the deep network,which has a certain reference value for the application of 3D scene restoration in related fields.