Research On Feature Extraction Algorithm Of 3D Model Based On Deep Learning

Three-dimensional shape models are becoming easier to obtain and capture more shape information.3D shape recognition has many applications in deep learning and computer graphics.The 3D model classification algorithm based on multi-view convolutional neural network(MVCNN)has achieved good classification results.CNNs are essentially limited to unknown geometric transformation,there lacks internal mechanisms in nature for handling the geometric transformations.The geometric transformation of the input data will affect the recognition rate of the algorithm.Based on Deformable Convolutional Networks(DCN)and MVCNN,a new deep convolutional neural network,De Conv-MVCNN,is designed and implemented on the Caffe deep learning framework.The main research contents of this thesis are as follows:(1)This thesis designs and implements a multi-view based 3D model classification network MVCNN on Caffe.The network has 9 layers,including 5 convolution layers,after each convolution layer,local response normalization and pooling operations are carried out.After five layers of convolution,a specific view pooling layer is designed,which is responsible for combining the information in multiple views of the 3D model into a single and compact shape descriptor,and the compact shape descriptor can extract the features of the three-dimensional model better,then three full-connection layers,and output the final classification results through the Softmax classifier.The average classification accuracy of multiple tests on the Model Net40 and Model Net10 reaches 89.9% and 90.5%,respectively.(2)To solve the unstable classification accuracy of traditional convolution neural network after geometric transformation of input data,this thesis designs and implements a deformable convolution network on Caffe platform based on the internal structure of convolution neural network.The network mainly includes two modules: deformable convolution and deformable ROI pooling.Compared with traditional convolution and pooling kernels,deformable convolution can effectively learn the offset of input data during the convolution process.The quantity greatly reduces the impact of geometric transformation of input data on subsequent networks,and improves the anti-geometric transformation ability of CNN.(3)For the problem of insufficient modeling ability of MVCNN geometric transformation,we proposed and implemented a new network model De Conv-MVCNN based on Caffe.The key idea is to add a deformable convolution layer which can enhance the ability of resisting geometric transformation between the input data and the first normal convolution.The deformable convolution layer first learns the offset from input data,then deformable convolution operation is performed,which significantly improves the robustness of the network.The results of several experiments on Model Net40 and Model Net10 show that the anti-geometric transformation ability of MVCNN is significantly improved.