Template-based Ear Modeling And Reconstruction

Human ear recognition,as a promising biometrics technology,has attracted more and more attention from domestic and foreign researchers.Compared with human ear images,3D human ear data contains more information.Using 3D human ear data for human ear recognition can get more robust result.Moreover,the application prospect of 3D reconstruction technology is very broad.Therefore,3D ear reconstruction has great significance in identity authentication,medicine,AR,VR.In the past,there have been few studies on 3D human ear reconstruction.The few 3D ear reconstruction methods rely on manual annotation of landmarks and manual intervention.The template-based ear modeling and reconstruction method is proposed in this paper for the first time.The main contributions and innovations are as follows:(1)A cascaded convolutional neural network is proposed to learn human ear feature representation.Aiming at the problem that the registration algorithm in the 3D ear reconstruction needs to provide initial corresponding point pairs,most of the current solutions are to manually mark the corresponding point pairs,that is,to mark two to four landmarks on the template model and the scan model.In order to improve efficiency and automation,a human ear detection network and a landmark positioning network is raised.A cascaded convolutional neural network can be used to learn the human ear feature representation,which can detect 55 ear landmarks without manual intervention.This way saves time greatly and makes preparation for the initial registration.(2)A template-based ear modeling and reconstruction algorithm is proposed and a mathematical model is established,which mainly includes two parts.The first part is to use the iterative closest point algorithm for rigid registration,and perform rigid transformation on the template model so that it is approximately in the same position as the target model in space.The standard iterative closest point algorithm is optimized,in which the adaptive threshold takes the place of the fixed threshold and the conditions that need to be met for calculating corresponding points are improved.Thus a more accurate rigid registration result is obtained.The second part is non-rigid registration.The non-rigid deformation is performed on the template model to eliminate the shape difference from the target model.This part is the key and difficult point.The traditional non-rigid registration algorithm is improved in this paper,which is used for the registration of human face and hand.After improving the algorithm,the algorithm can get good results on the 3D ear reconstruction.Firstly,the deformation graph on the template model is constructed and the distance calculation method is optimized,so that the sampling vertexs of the deformation graph are evenly distributed on the template model,and the constructed deformation graph is more accurate.Secondly,the energy objective function for the human ear is designed.A smoothing term is added to the objective function,which enables the template mesh to deform smoothly.After continuous optimization of the non-rigid registration algorithm,it is more suitable for 3D modeling of the human ear and its performance improves greatly.(3)Finally,the application of 3D ear reconstruction is studied.After reconstructing the three-dimensional ear models on the human ear dataset,an improved two-step iterative nearest point algorithm was used to perform the identification experiment.To solve the problem that the iterative nearest point algorithm is prone to fall into the local minimum,the elite retention strategy is used to reduce the search range and accelerate convergence.In order to improve the efficiency and accuracy of the nearest point pairs calculation,we use an index structure and add unique constraints.The experiment obtains a high recognition rate and has verified the feasibility and effectiveness of using the 3D human ear model for identity authentication.