Research On Deep Learning Based Multi-view 3D Object Classification

3D object classification is an important research topic in the field of computer vision.Good classification performance can support subsequent advanced processing.The traditional classification method extracts data features manually.The disadvantage is that manual setting is time-consuming and labor-intensive and dependent on subjective experience.The deep learning model can automatically learn the deep features of the data,reducing the subjectivity of manual classification,and the accuracy is relatively high.Multi-view representation is one of the important representations of 3D models.It conforms to the human eye's visual perception of the object and can be used as input data for the deep learning model.This paper focuses on multi-view 3D object classification problem based on deep learning,the method of multi-view representation and selection of 3D models is studied.The deep learning model is used to classify 3D object,aiming at improving the accuracy and further improving the calculation speed.The main work of this paper is as follows:(1)Aiming at the problem of the multi-view view data lacks the related information between views,this paper proposes a multi-view and panoramic view representation method for the 3D model.In this paper,the multi-view view data of the 3D model is constructed firstly.Based on this,the panoramic view data is further constructed,the related information between the views is added,and the two are combined to form a complete multi-view dataset.Complete information of the model solves the problem of missing information in multi-views,experiments show the effectiveness of the model representation method.(2)Aiming at the problem of multi-view data samples and a large number of redundant fuzzy features,this paper proposes a multi-view saliency analysis and selection algorithm based on the constructed multi-view dataset.In this paper,the saliency of multi-views is analyzed by gradient calculation.The views are sorted according to the gradient.On this basis,the views with significant features are selected to solve the problem of low computational efficiency.The Experimental results on public dataset show the effectiveness of the multi-view analysis and selection algorithm.(3)Aiming at the problem of less feature information for single view and insufficient feature description in multi-views,Based on multi-view data construction,analysis and selection,this paper proposes a multi-view classification model based on aggregate feature description.In this paper,the multi-views are first pooled,the feature information of multiple views is aggregated,and the aggregate descriptor representing the 3D model is generated.Based on this,the VGG-M model is used to classify the multi-views on the public dataset.The experimental results show that the classification model based on aggregated feature description has a better classification effect.