3D Model Recognition And Retrieval Based On Deep Learning And Multi Shape Descriptors Fusion

With the rapid development of information technology and artificial intelligence,the data processing ability of computers has been significantly improved,which supports the transformation from two-dimensional images to three-dimensional models.3D modeling finds its place in many fields,i.e.,in the chemical industries,the real-time monitoring and disaster warning can be realized by the automatic 3D scene modeling.Current 3D modeling software usually has the advantages of high efficiency and low economic cost.3D modeling greatly saves the manpower,while the generated 3D models suffer from the partial missing and low resolution,which sets up obstacles in practical application.This paper studies the algorithm of recognizing and retrieving 3D models to automatically replace the low-quality models with the high-resolution models in 3D scene,which contains abundant and profound researching significance.3D model recognition and retrieval are the most difficult and important problems in 3D series.The point cloud has the characteristics of high dimensionality and disorder,so any single feature cannot accurately depict the model shape.In this paper,we calculate the scale-invariant heat kernel signatures as local features based on the point cloud and meshes.Then we use the non-parametric kernel density estimation to forecast the probability density distributions of local features.Due to the inevitable error of kernel density estimation,we combine the auto-encoder and multilayer perceptron to form a deep functionally bifurcated semi-supervised coding network to extract the high-level features of the probability density distributions,and take the outputs of encoder as the shape descriptors.In order to completely eliminate the error caused by kernel density estimation,we furthermore cascade the scale-invariant heat kernel signatures to the spatial-transformed point cloud matrix to propose a multi-input and single-output convolutional network based on Point Net,which can achieve a higher retrieval accuracy,compared to the former method.Finally,we calculate the weights of the former two shape descriptors based on the information entropy of the first tier criterion to fuse the two shape descriptors.The experimental results demonstrate the shape descriptor after fusion has stronger discrimination ability.In this paper,we also involve the random sampling of point cloud and the plurality-based voting strategy.The former diversifies the training samples and strengthens the robustness of network,while the later can better fuse the class information into shape descriptors.