Research on Multi-View Image Reconstruction Technology Based on Auto-Encoding Learning

Abstract

Traditional three-dimensional (3D) image reconstruction method, which highly dependent on the environment and has poor reconstruction effect, is easy to lead to mismatch and poor real-time performance. The accuracy of feature extraction from multiple images affects the reliability and real-time performance of 3D reconstruction technology. To solve the problem, a multi-view image 3D reconstruction algorithm based on self-encoding convolutional neural network is proposed in this paper. The algorithm first extracts the feature information of multiple two-dimensional (2D) images based on scale and rotation invariance parameters of Scale-invariant feature transform (SIFT) operator. Secondly, self-encoding learning neural network is introduced into the feature refinement process to take full advantage of its feature extraction ability. Then, Fish-Net is used to replace the U-Net structure inside the self-encoding network to improve gradient propagation between U-Net structures, and Generative Adversarial Networks (GAN) loss function is used to replace mean square error (MSE) to better express image features, discarding useless features to obtain effective image features. Finally, an incremental structure from motion (SFM) algorithm is performed to calculate rotation matrix and translation vector of the camera, and the feature points are triangulated to obtain a sparse spatial point cloud, and meshlab software is used to display the results. Simulation experiments show that compared with the traditional method, the image feature extraction method proposed in this paper can significantly improve the rendering effect of 3D point cloud, with an accuracy rate of 92.5% and a reconstruction complete rate of 83.6%.