Research On Non-Rigid 3D Model Retrieval Algorithm Based On Deep Learning

3D model is more and more widely used in life.However,it takes a lot of time and energy to reconstruct complex 3D model.Sometimes,only local modification is needed on the existing model to get the desired model.Therefore,3D model retrieval has become a research hotspot.Effective and efficient features are the key to the application in the field of 3D retrieval.Deep learning has been proved to be very effective in various visual applications such as image classification and target detection.However,it has not been successfully applied to 3D shape recognition.Because it is impossible to train the 3D model directly into the neural network,most of the methods based on in-depth learning now extract the features manually or process the 3D model into structured data that can be input into the neural network to extract the features.In this thesis,a method of 3D feature extraction based on different modal data is proposed.Firstly,the coordinates and scales of the 3D model are normalized,and two kinds of data(voxel representation and pixel representation)are used to deal with the problem of object recognition.Then a convolutional neural network is constructed,which is used as the input of the network to extract visual and geometric information,and to improve the recognition ability of single modal features by using in-depth learning.The geometric descriptors and view descriptors of the 3D model are extracted respectively based on convolutional neural network,and the two feature descriptors are fused into a network to find the relationship between the patterns and form the fused feature descriptors,which are then applied to the classification and retrieval of the three-dimensional features.The multi-feature fusion layer can not only learn more discriminant features of the two descriptors,but also supplement the relevant information between the two descriptors.Compared with using these two representations alone,the classifier generated by this method is much better and the retrieval efficiency is also improved.The experimental results of non-rigid model datasets based on multiple benchmarks show that the fused features not only contain information about the surface of the model,but also describe the internal properties of the model.Compared with traditional descriptors and single feature descriptors based on deep learning,fusion descriptors based on deep learning have more advantages in classification and retrieval.