Non-rigid body three-dimensional shape fan-shaped convolution feature extraction method and system

Abstract

The invention discloses a non-rigid body three-dimensional shape fan-shaped convolution feature extraction method and system, and the method comprises the steps: S1, extracting a salient point set of a non-rigid body three-dimensional shape based on the farthest geodesic distance sampling; S2, performing feature extraction on the salient point set of the non-rigid body three-dimensional shape through a fan-shaped convolutional neural network to obtain a depth intrinsic feature of the non-rigid body three-dimensional shape; wherein the fan-shaped convolutional neural network has four layers, the first layer is three-dimensional fan-shaped convolution, the second layer to the fourth layer are two-dimensional fan-shaped convolution, and output channels are 256, 512 and 1024 respectively; and S3, through feature splicing and maximum pooling, processing the obtained depth intrinsic features of the non-rigid body three-dimensional shape so as to obtain the depth global features of the non-rigid body three-dimensional shape. According to the method, the point cloud scale of the non-rigid body three-dimensional shape is reduced, the subsequent calculation loss is reduced under the condition that the point cloud characteristics are kept, and the deep intrinsic characteristics with higher expression ability can be extracted.

Description

technical field [0001] The present invention relates to the technical field of deep learning and computer vision, in particular to a fan-shaped convolution feature extraction method and system for a non-rigid three-dimensional shape. Background technique [0002] With the rapid development of "Internet +" and computer technology, the 3D shape information in various fields has increased massively, and various 3D shape libraries have been formed. Three-dimensional shapes are divided into rigid bodies and non-rigid bodies. For the classification and retrieval of rigid three-dimensional shapes, there have been relatively mature researches and good results have been achieved. Due to the huge difference between non-rigid body and rigid body, especially the 3D shape of non-rigid body has a hinge structure, there are many types and complex forms of positional relationship transformation of points inside the shape, if the more mature scheme on rigid body is directly applied to non-ri...

AI Technical Summary

Problems solved by technology

Due to the huge difference between non-rigid body and rigid body, especially the 3D shape of non-rigid body has a hinge structure, there are many types and complex forms of positional relationship transformation of points inside the shape, if the more mature scheme on rigid body is directly applied to non-rigid body, the effect is often poor

[0003]In some existing non-rigid body feature extraction methods, the features based on artificial design have limitations, such as poor applicability, sensitivity to noise, large amount of calculation, However, the method based on visual codebook can realize automatic learning, but its ability to obtain high-level semantic features is limited

The feature extraction method based on deep learning has certain limitations due to the different learning data sources: the method based on artificial features relies on manual selection of parameters; the method based on projected images ignores the local details of the shape and the global structure during the transformation process. The mutual relationship reduces the feature discrimination; based on the three-dimensional voxel method, the shape information contained in the voxelized binary expression is limited, and the calculation complexity for high resolution is large, and it limits the feature recognition for low resolution. Discrimination; based on the original data method, it is difficult to design a deep learning network that can adapt to the characteristics of the original three-dimensional data

Images