Research On 3D Model Classification Algorithm Based On Multi-local Salient Views

As the continuous development of 3D scanning,modeling and computer vision technology,the 3D model has become a new type of multimedia data,which is widely studied and applied in the fields of entertainment,medicine,industry,etc.How to effectively retrieve and reuse these 3D models has become an urgent concern and research focus in this fields.Traditional 3D model classification methods are mainly based on artificially designed feature descriptors to achieve classification and retrieval.Due to the continuous increase in the complexity and number of 3D models,when using these methods it is difficult to capture the essential characteristics of 3D models objectively and reasonably,which further effects classification accuracy.Deep learning can learn the multi-level abstract representation of complex data automatically,and it has achieved remarkable results in many applications,such as image recognition and speech recognition.It provides a new way for the effective classification of 3D models.Aiming at the defects of the traditional 3D model classification methods,using deep learning technology,we propose a 3D model classification algorithm based on multi-local salient views.The main work of this thesis includes the following aspects:(1)A multi-local salient views representation model for the 3D model is proposed.We design and implement an automatic construction method for multi-local salient view set,so that the constructed multi-local salient view set not only conforms to the input data type of the deep learning model,but also represents the three-dimensional model well.(2)An ensemble convolutional neural network model is proposed.With the multi-local salient view as input,we extract the feature for each local salient view using Caffe Net,and classify the 3D models using voting integration strategy,so as to enhance the learning ability and classification ability of the network.(3)Based on the above research,we design and implement the proposedalgorithm.The classification accuracy rates of this algorithm on the Model Net10 and Model Net40 benchmarks are 94.5% and 88.625%,respectively.It demonstrates the feasibility and validity of our algorithm.