3D Shape Recognition Based On 2D-CNN Adaptive Feature

In recent years,deep learning technology in the field of 3D model retrieval has achieved rapid development.With the characteristics of complex structure and large amount of information,3D shape converted into 2D images by the recognition method has achieved good performances.Both the projection method and the feature extraction method of deep learning have different effects on 3D shape recognition.We focus on the feature recognition of 2D images of 3D shapes.The specific works are as follows:(1)The principal component analysis method(PCA)has non-correlated feature compression capabilities and combines it into deep learning to enhance the feature extraction capability.The regionalization method is used to greatly focus on the expression of local features.Finally,mean fusion of multi-view 3D shapes represents the global feature descriptors of the 3D shape.After adding this method in the CNN framework,a trainable balanced principal component network(BPCN)is formed.The classification result and mAP value on ModelNet40 dataset are increased from 93.8% and 93.2% to 95% and 94.6%,respectively.The average retrieval accuracy of top-10 is 98.5% for 50 shapes in each class and 20 shapes for each class in the database.The average retrieval accuracy of top-10 on ModelNet10 is 100%.(2)The combination of principal component analysis and deep learning is optimized based on BPCN,and the improved box-cox called adaptive power weight is used to weight the singular values,which greatly reduces the number of parameters and accelerates the fitting of weights through the parameters of the neural network.The structure is divided into three parts.First,the features are divided into sub-regions to form a series of local features.Second,the maximum or average singular values of these features are calculated as the feature descriptors of the 3D shape.Parameter learning of singular value features through back propagation.This method combines with CNN to form an adaptive dynamic singular value network(DSVN).The classification result and mAP value on the ModelNet40 dataset are 95% and 94%,respectively.The average retrieval accuracy of top-10 is 98%,and the average retrieval accuracy of top-10 on ModelNet10 is 99.9%.DSVN has reduced the training time by one-third compared with BPCN.(3)Aiming at the problems of multi-angle 3D model projection and feature classification,an tilted projection method and a shape library classification method are proposed.The tilted projection method projects within the "effective perspective" of the 3D shape to increase the difference of each shape,which has a good result on rigid shapes.The shape library classification method establishes a shape library adaptively based on the training set.By comparing the features in the shape library during the classification process,this method maximizes the fault tolerance of the features and removes the misclassification results in the CNN.The tilted projection method contributes 0.4% to 1.3% based on the classification accuracy of 95%.Shape library classification improves 0.5% to 1.5% compared to SVM classification method on the same basis.The highest classification accuracy on ModelNet40 dataset is 96.9%,and the highest accuracy classification on ModelNet10 dataset is 98.5%.The algorithm running time is reduced by 50% to 66% compared to SVM.