Multi-view reconstruction method based on deep learning contour network

Abstract

The invention provides a multi-view reconstruction method based on a deep learning contour network. The method comprises steps of introducing a deep learning architecture, carrying out three-dimensional shape coding, constructing a contour network, and carrying out network training and testing. To be specific, a deep learning contour network is introduced and the used for learning three-dimensional shape codes of one or more input images; a new view is generated by using a code adjustment decoder; with introduction of a contour-based proxy loss, when the decoder does not include three-dimensional representation, the network uses the two-dimensional loss to encode the three-dimensional shape, wherein the two-dimensional loss is not limited by the three-dimensional representation resolution;and a massive spotted object data set network is generated and pretrained and then micro adjustment of the contour network of the data set is carried out. According to the invention, the three-dimensional shape is learned by using the neural network and the contour generated in the new view forces the network to code a three-dimensional shape; information of multiple views is combined; and the multi-view reconstruction performance is improved.

Description

technical field [0001] The present invention relates to the field of view reconstruction, in particular to a multi-view reconstruction method based on deep learning contour network. Background technique [0002] Multi-view reconstruction is a method to restore the 3D model of the scene by using multiple pictures of different perspectives of a scene. The multi-view of natural scenes Figure three Three-dimensional reconstruction has always been a basic problem in the field of computer vision. By reconstructing the three-dimensional model of the target, the target can be quantitatively analyzed and the relevant information of the target can be processed. Multi-view reconstruction technology is applied to medical imaging. Through view reconstruction, on the one hand, biomedical imaging technology can obtain visualized structural or functional information for biological research or clinical diagnosis; on the other hand, human body structure can also be obtained through reconstruc...

AI Technical Summary

Problems solved by technology

Although there is a lot of research on multi-view reconstruction, there are still some challenges in combining information from multiple views and improving the reconstruction performance accordingly when the original view images are few, the baseline is wide and the lighting conditions are complex.

Images