The Research And Application Of 3D Shape Retrieval Based On Sketches

With the maturity of the 3D scanning equipment and the development of related 3D modelling technology,it has become easier to obtain and produce 3D models.There are large amounts of 3D models being produced everyday world wide.The range of applications for 3D models also has become more and more extensive.The vigorous development of a variety of 3D modelling technologies and printing technologies have also brought a large number of publicly available 3D model databases.Three-dimensional modeling is complex,and the cost of three-dimensional modelling is high,3D model retrieval technology can help users in the circumstances of information overload efficiently access to target information.Compared with other methods,the 3D model based on hand-drawn sketch retrieval has the advantages of being simple and convenient.Two-dimensional hand-drawn sketches and 3D models are not consistent in the data dimension,and there is no direct similarity between the two.In order to solve the difficulties of the 3D model retrieval method based on sketches,this thesis proposes a retrieval scheme based on the Improved Siamese Convolution Neural Network.The main work of this thesis is as follows: First of all,for the reason of the sketch and the 3D model are difficult to match directly,we need to use dimensionality reduction to projection the 3D model and generate two-dimensional projection atlas,which transforms the matching problem of sketch and 3D model.Secondly,the current state-of-the-art technology is almost always calculated using a large number of "best views" for 3D models,and it is desirable that the query sketch matches a two-dimensional projection in a 3D model using predefined features.Because the "best views" are subjective and ambiguous,this makes the matching input blurry.This article suggests using a simpler way to define a view,dramatically reducing the number of views by learning cross-domain similarity and learning the features of sketches and views rather than relying on the "best views" and feature extraction of hand-drawn sketches.Thirdly,in order to solve the problem of cross-domain matching,two Siamese Convolutional Neural Networks(CNNs)are used to propose the cross-domain matching using Improved Siamese Convolutional Neural Networks,and the similarity between domain and cross domain was solved.Fourthly,In this thesis,we add a softmax loss on the basis of the contrastive loss function of the original Siamese Network,and then define a new loss function to "align" the results of the two CNN models,which makes the feature's identification significantly enhanced.Experimental results on three benchmark databases demonstrate that the proposed method offers a better performance when compared with several state-of-the-art approaches.