3D Digital Image Virtual Scene Reconstruction Algorithm Based on Machine Learning

Abstract

3D modeling has been widely used in industrial, medical, military and other fields. Trying to solve the issue that the traditional 3D reconstruction model is ineffective in processing digital image feature extraction in virtual scenes, this study adopts a multi-view stereo vision algorithm and neural network to optimize it based on the traditional 3D reconstruction algorithm. Then, the spatial attention mechanism and the channel attention mechanism were combined to generate a Convolutional Block Attention Module (CBAM) model, and the CBAM was used in a multi-view stereo vision algorithm model. The model's performance is tested, and it is found that the convergence speed is faster in training, the loss function value is lower, and the overall model's performance is better. In the test, compared with the other three models, the accuracy of the proposed model is improved by 17%, 9% and 3% on average. The integrity of MVSNet-CBAM was enhanced by 28%, 14% and 9%, respectively. The experiment verifies the validity, which aims to provide a reference for 3D digital image virtual scene reconstruction.