Research On 3D Reconstruction Method In Images Based On Deep Learning

Image-based 3D reconstruction has always been an important research direction in the field of computer vision.Traditional methods often require more prior knowledge to support the completion of reconstruction tasks,and some methods have poor generalization ability and cannot reconstruct images of wide baseline images.In recent years,with the continuous development of deep learning,the image reconstruction method based on deep learning has achieved better results and requires less prior conditions.This paper systematically analyzes and summarizes the research foundation of 3D reconstruction based on deep learning.Based on the network structure of deep learning method 3D-R2N2,an improved method and a new network structure are proposed for the problem of poor reconstruction accuracy.The effectiveness of the method is verified by experiments.The main work of this paper is as follows:(1)This paper proposes two improved Inception-resnet modules and a reconstruction network based on the improved Encoder module.By adding a modified Inception-resnet module to the Encoder module,the fully connected layer is removed,and the global average pooling is used to enhance the network's ability to extract detailed features while reducing network parameters and computation.The experimental results demonstrate the superiority of the method in terms of time consumption and reconstruction accuracy.(2)The differences between the reconstruction effects of several reconstruction networks using different feature extraction modules on different objects are summarized by experiments.Based on this,a multi-feature reconstruction network based on single image is proposed.The network utilizes several different feature extraction modules to extract multiple features and aggregate them in the 3D-LSTM structure.In this way,the feature extraction capability of the network for a single image is enhanced,thereby improving the reconstruction effect.The experimental results show that the average reconstruction accuracy of the proposed method is significantly improved in the reconstruction of objects in a single image,and in the reconstruction of 13 types of objects,the reconstruction of objects such as lamps,chairs and ships is better.(3)This paper proposes a multi-feature reconstruction network based on multiple images combined with Attention mechanism.It extracts multiple features through multiple feature extraction modules to enhance the feature extraction effect.At the same time,3D-LSTM-Attention can help the network cognitive features to achieve a more targeted reconstruction and improve reconstruction effect.After experiment,the method has higher average IoU value,lower cross entropy loss and better reconstruction effect on multiple images.