Research On Sketch-based 3D Model Retrieval

With the rapid development of computer vision technology and digital multimedia technology,3D model processing technology has been used more and more widely in the fields of film and television animation,medical research,computer simulation and industrial product design.Currently the use of a platform or tool to retrieve the target 3D model quickly on the Internet or database has become an urgent demand in many research areas.It is particularly important to retrieve the target 3D model accurately from the database via free expression.At present,there are still some improvements in the retrieval effect of 3D model retrieval technology based on freehand sketch,and the difficulties include two aspects:Flexible expression and accurate identification of the searcher's intention;the comprehensiveness and accuracy of 3D model feature extraction.Therefore,in the research of 3D model retrieval technology based on freehand sketch,this paper focuses on the sketch recognition algorithmand the feature extraction algorithm of 3D model.The main research work of this paper is as follows:(1)A new stroke recognition algorithm is proposed.The stroke recognition algorithm consistsof two parts,the stroke fusion algorithm and the stroke segmentation algorithm.With the help of the stroke fusion algorithm,the problem of the input stroke with inconsistent direction and the inaccuracy expression could be averted.Using the stroke segmentation algorithm,the graph is divided into basic elements for identification.Based on that,the basic graphic elements of the sketch are identified.The basic graphic elements represent the projection contours inside the object,representing the local features of the object.A sketch should not be only considered as the shape of the external contour,but also the relation between the basic graphic elements.The experimental results show that the combination of the stroke recognition algorithm and the basic primitive recognition method effectively improves the accuracy of the two-dimensional query sketch recognition.(2)An improved algorithm of 3D model for obtaining 2D projection images is proposed.The algorithm consists of two steps:Firstly,generate the projection image.A projection algorithm based on the virtual photographing with the longitude and latitude of the sphere is adopted.This method divides the longitude and latitude of a sphere at interval of 200°,thus 162 projection images are produced through the way of virtual photographing.Secondly,choose the projection image.An optimized proportion algorithm is proposed,which is used to select the optimal 2D projection image.In the 162 projection images,the 2D projection images which are inaccurate in matching with the sketch would be eliminated.The experimental results show that the projection algorithm based on the virtual image of sphere longitude and latitude improves the retrieval accuracy.At the same time,the optimized proportion algorithm improves the retrieval efficiency.(3)The feature extraction algorithm based on feature descriptor is proposed,which combines the advantages of Fourier descriptors,Zernike moment,roundness and eccentricity.The four feature descriptors are weighted combination to represent the global features of the 2D projection image and the freehand sketch.Meanwhile,this paper adopts the feature extraction algorithm based on equal point sampling and feature extraction algorithm based on polygon to extract the global feature of the external contour.Via the feature extraction algorithm based on equal point sampling,on the one hand the translation and rotation targets do not move on the contour line,on the other hand,the translation and rotation do not affect the distance from each point of the contour line to the geometric center of the shape.The feature extraction algorithm based on polygon regards the larger curvature on the external contour as characteristic points of a curve,which makes the fitting of the characteristic points of the curve more detailed,as much as possible to remain the image feature information.The experimental results show that the proposed global feature extraction algorithm can effectively improve the recall and precision of 3D model retrieval based on freehand sketch.(4)A PCA-DAISY local feature descriptor is proposed to extract the local features of the image,which can represent a large amount of valid information in the interior area of the object.Firstly,the DAISY descriptor is composed of a set of vectors of convolution directional diagram derived from the Gauss filter.Principal component analysis theory is used to reduce the dimensionality of DAISY descriptors,and also to reduce the error rate.Furtherly,we use the Fisher encoding algorithm to quantize the PCA-DAISY descriptor,which forms a retrieval method of "probabilistic visual vocabulary".The experimental results show that the retrieval efficiency of PCA-DAISY descriptor does better than that of other descriptors,and the Fisher encoding algorithm provides with higher retrieval accuracy.(5)The local feature extraction algorithm based on skeleton is proposed.The algorithm consists of two steps:Firstly,generate the skeleton of the image.The advanced skeleton strength map algorithm is proposed,which select the key points from the skeleton strength map,and use the Kruskal's minimum spanning tree algorithm to connect them to generate the skeleton of the image.Then,compare the skeleton similarity between freehand sketch and 2D projection image.A new skeleton feature matching algorithm based on histogram is proposed to research on the feature matrix of the histogram,and the minimum value of each row in the matrix is chosen to calculate the similarity distance between the two skeletons.The experimental results show that the local feature extraction algorithm based on skeleton can generate skeleton branches of important visual components and provide with higher retrieval accuracy.