Multi-view Deep Panorama For 3-D Shape Recognition

The popularization of artificial intelligence has aroused the trend of using machine learning methods to do research in various fields,especially in the field of image processing.This is because two-dimensional images are structured data,easy to acquire a large amount of data sets,and can be fully used in the training and learning of models.It provide a great convenience to researchers.With the rapid development of computer technology and Internet,multimedia data used by people begin to transition from traditional two-dimensional image data to three-dimensional data.The three-dimensional data structure is complex(unstructured data),the early data acquisition was difficult and the amount of data is relatively small.There were some difficulties in using the machine learning method to study and the development is relatively late.However,the three-dimensional data contains richer information.And with the development of 3D scanning technology and equipment,the amount of data is also gradually increasing,and the fields of application are more and more widespread.The related research on retrieval and classification is of particular concern.Through in depth study of three-dimensional shape recognition algorithm,summarizes the existing research methods,the paper propose a new method for classification and retrieval of three-dimensional model:(1)According to the shape representation and classification of 3D models,this paper introduces a concept of multi-view panorama as the descriptor of threedimensional shape,and uses a set of panorama images to describe the position and orientation of the object surface in three-dimensional space.And use multi-view neural network to do three-dimensional shape classification.In this method,the panoramic image is obtained by cylindrical projection of the three-dimensional model.The multi-view panoramic image information can be aggregated into a more compact three-dimensional shape descriptor by using the multi-view convolutional neural network framework.Finally,we can get the classification probability of threedimensional model.The 3D classification method based on multi-view depth panorama effectively solves the problem of shape representation of 3D models,and the classification effect has obvious advantages.(2)In order to solve the problem of 3D model retrieval,this paper combines thetwo-dimensional image retrieval method with multi-view panorama to train and finetune the multi-view convolution neural network framework.Extracted six kinds of eigenvectors of the panorama from the pre-trained and fine-tuned CNN model.And use the Euclidean distance estimated the distance between the panoramic images.Experiments show that the retrieval performance of this method outperforms other methods.After the fine-tuning training model,the Mean Average Precision(MAP)of retrieval is significantly higher than that of the pre-training model.