3D Reconstruction Based On Single Flower Image

In recent years,with the rapid development of 3D display technology,3D Flower model has a very important application in scene design,computer games and so on.The 3D reconstruction of plants has always been a key and difficult point in the field of computer graphics and computer vision,and the modeling process of flowers is more complex and difficult because of its geometric structure and variety.How to acquire the three-dimensional data of flowers quickly and effectively has become the main bottleneck of its further development.Traditional three-dimensional reconstruction techniques,such as Structure from motion method and Photometric Stereo method,need to recover the three-dimensional structure of objects from two or more pictures with different perspectives.Although the 3D reconstruction based on single image has the advantages of simple input and convenient data acquisition,it also has some problems such as the lack of information of the occluded perspective.With the development of deep learning in recent years,it is possible to infer other perspective pictures from a single perspective picture.In this thesis,we will make full use of the symmetry of flower ecological structure and deep learning method to make up for the lack of perspective information,so as to realize the flower reconstruction of a single image conveniently and quickly.The main research contents and conclusions of the thesis are as follows:(1)Construct a data set of three-dimensional flower modelsIn order to solve the problem of the lack of flower 3D model data set in the existing 3D reconstruction research,this study used 3ds Max tool to segment the flowers in the 3D scene model collected on the network,a total of 4106 separate flower models were extracted.After preprocessing each flower model,Panda3 D was used to render it,and RGB images and depth maps of each flower model from different perspectives were obtained.A flower data set that could be used to input deep neural networks was constructed.(2)Generation of flower multi-view RGBD image based on single flower imageIn order to solve the problem of self occlusion and missing depth information in the existing algorithm of 3D reconstruction of flowers based on a single picture,this thesis applies an algorithm of generating depth information of flowers from different perspectives and different perspectives based on convolution neural network.Input single view RGB flower image and target view information,and generate multiple different view images and depth maps through Encoder-Decoder network structure.The experimental results show that for flowers with a simple structure and strong symmetry,the mutual information between the depth map generated by this method and the real depth image is 0.6743,which basically infers the occluded part,which better reflects the structure of the flower With contours.(3)Mesh 3D model reconstruction after generating multi-view image based on RGBD imagesAiming at the problem that the camera poses of the multi-view RGBD pictures generated in(2)are not uniform,this thesis studies the ICP registration method based on point cloud to register the point cloud of different angles,so as to fill the missing part of point cloud.After getting the registration point cloud,the greedy projection triangulation algorithm is used to generate the mesh.Experimental results show that the three-dimensional mesh model reconstructed by the algorithm in this thesis can better maintain the global characteristics of the flower model,the reconstruction accuracy is acceptable,The average chamfer distance from the point cloud of the real flower model is 0.2703,and can faithfully reflect the target flower structure.