Research On Sketch-based 3D Model Retrieval And Its Implementation

Three-dimensional model have been widely used in medical,education,industrial design,virtual reality,enhanced reality and other fields.With the development of various 3D modeling techniques,there have been a large number of publicly available 3D model databases.As the cost of three-dimensional modeling being high,so using information retrieval technology to reuse existing three-dimensional model to build three-dimensional scene has a great demand.Three-dimensional model retrieval technology has become a hot topic in computer graphics.Hand-painted sketches is universally available skill,it has unique advantages as a new method of human-computer interaction.It is of great research value to use hand-drawn sketch to retrieve three-dimensional model compared with other methods with simple and convenient advantages.However,the abstraction of hand-drawn sketches and the complexity of 3D models make it difficult to retrieve 3D models based on hand-drawn sketches.In order to solve the difficulties and challenges in the three-dimensional model retrieval method based on hand-drawn sketches,this paper focuses on how to solve the three-dimensional model retrieval based on hand-drawn sketches by deep learning technology,and integrates a set of deep convolution neural networks as the core Three-dimensional model retrieval scheme.This paper mainly completes the following work:First,the problem of hand-drawn sketches and 3D models is difficult to match directly,the 3D model is rendered as a line projection map similar to the hand-drawn sketch,which transforms the problem into the matching problem of the hand-drawn sketch and 3D model projection graph.Secondly,the feature expression ability of hand-crafted feature extraction method is not strong and the generalization ability is weak.The feature extraction method of deep convolution neural network is studied,and the advanced abstract features with strong expression ability and generalization ability are constructed.Thirdly,aiming at the problem that the hand-drawn sketch and the 3D model projection are different in style,it is difficult to directly measure the degree of similarity.It is proposed to learning a similarity metric to measure the similarity between the hand-drawn sketch and the 3D model projection.After embeding the sketches and the projection images of 3D model into the same metric space,we effectively solved the problem of cross-domain matching.Fourthly,in order to make our algorithm invariant to various variants of sketches,the constraints are put into the neural network model,so that each class of samples in the learned metric space has the same kind of high degree of cohesion and low coupling between different classes.This enhances the robustness of the retrieval algorithm.Fifthly,in order to meet the high availability requirement of the retrieval system,an auto-encoder is designed to reduce the dimensionality of the feature vectors.The computational efficiency is effectively improved without loss of the retrieval precision,so that the scheme can be extended to the large-scale database use.At the end of this paper,we experiment with multiple benchmark databases,and prove the advanced nature of the algorithm by comparing with the existing state-of-art methods.