3D Model Retrieval Based On Deep Convolution Neural Network

With the rapid growth of 3D models and their wide application in various fields,an urgent problem is to develop a search engine to help users quickly and easily find the expected 3D models.Multi-view convolutional networks can effectively improve feature discriminability of 3D models through deep learning.Moreover,it can support sketch interface based 3D retrieval.Therefore,3D model multi-view convolutional networks have attracted more and more research interests.This thesis is focusing on this research topic and the main work is as follows:1.The related feature extraction approaches in 3D model retrieval are summarized and analyzed.The related research is classified into 2D shape features,3D shape features and multi-view features.At the same time,the theoretical framework of convolutional network is given.2.On the basis of the existing residual convolution neural network,the weighted loss function is proposed to improve the discriminability of the view features of 3D models.Firstly,the 3D model is rendered to obtain different views.Then,a residual network expansion module is used to increase depth of the network.Meanwhile,the weighted loss function is defined by combining the center loss function and the cross entropy loss function.As a result,it can solve the problem that the intra-class distance is less than the inter-class distance.Experiments on ModelNet datasets show that the algorithm's performance is excellent in 3D model classification.3.A convolutional Network Cascade is proposed to improve the retrieving accuracy of sketch based 3D retrieval.The 2D sketch is one of the most simple and efficient interface to search 3D shapes.However,since the visual difference is be in 2D sketch and 2D views of 3D model,the retrieval accuracy is not high as other interface,like 3D model examples.In this paper,a convolutional Network Cascade is proposed to consider not only the separability of features,but also the approximation errors of sketches and views.The proposed network structure consists of two parts.The first part is to generate a network to improve the similarity of visual features between rendered view and sketch.The second part is residual network,which is used to learn the inter-class separability of convolution features.Meanwhile,the weighted objective function is used in the objective function.In the experimental analysis,the paper validates the proposed structure by SHREC'13 and SHREC'14 benchmark data sets.It can be found that this convolution network Cascade improves the clustering of features and greatly improves the retrieval accuracy,which exceeds the retrieval accuracy of manifold embedding algorithm.