Research And Implementation Of Robust 3D Reconstruction Algorithm Based On Intergranular Relationship Reasoning And Information Fusion

With the increasing demand in the fields of Augmented Reality(AR),autonomous driving,Virtual Reality(VR),3D printing and 3D modeling,the research value and significance of robust single-view 3D reconstruction algorithm in the field of image understanding and computer vision are reflected.Most of the existing single-view 3D reconstruction methods focus on the reconstruction of a smooth and complete global shape while ignoring the refined fine-grained details.This leads to the reconstruction of objects that has a good overall shape but lacks some local fine-grained geometric details.In addition,the existing single-view 3D reconstruction methods are difficult to accurately reconstruct 3D shapes with complex topological structures,which leads to problems such as missing shapes and blurred details of reconstructed objects in the 3D space domain.Therefore,how to accurately recover three-dimensional shapes with high fidelity and rich details from single-view images is a challenging problem in 3D reconstruction tasks.In this paper,we propose a robust model based on deep learning to solve the above problems,which can make the reconstructed object have fine local fine-grained details while maintaining a good global shape to achieve high-fidelity and rich details 3D reconstruction.The main work of this paper is as follows:1.Through the analysis of a large number of 3D synthesis and real image datasets,we find that reconstructed objects usually have richer fine-grained information near the edges and corners.In order to better capture and use the edge detail information in the input image,we use a separate Difference of Gaussians(DoG)batch network to extract the edge information at the edges and corner from the input image to capture the complex topological structure of the object in the 3D space and then use the several side branches to extract the discriminative features at multiple scales from the input image.Finally,we dynamically fuse the prediction results of all branches to obtain the final 3D shape.2.In order to better reconstruct objects at different scales and refine the edge information of fine-grained objects at different scales,this paper designs the Multi-Scale Feature Interaction Block(MFIB)and Multi-Scale Gaussian Difference Block(MGDB).The multi-scale feature interaction block(MFIB)uses feature interactions between different scales to acquire more local fine-grained details while learning global shape information.The multi-scale Gaussian difference block(MGDB)uses DoG operations on input images of different scales to extract edge geometry and comer information at different scales.Finally,this paper dynamically mixes the two parts of the prediction information to get the final 3D shape.3.In order to solve the problem that the previous methods ignore the refined fine-grained detail information,we carefully design an accurate multi-granular learning module to achieve robust feature representation and refined 3D reconstruction.Compared with the previous single-view 3D reconstruction methods,our method not only captures the most discriminative features in the original input image but also extracts detailed features at different granularities.In addition,we design a cycle loss function between multiple granularities to enhance the consistency and inter-granular correlation.4.we design a 3D reconstruction system which applies our algorithms on the Web platform.In addition,users can choose the most suitable 3D reconstruction algorithm for application from the three algorithms integrated in the system according to different task goals and requirements.