The Research Of 3D Model Retrieval Algorithm Based On Multi-view Graph Neural Network

With the rapid development of 3D modeling technology,an increasing number of 3D models are stored in various fields.For example,three-dimensional models are widely used in industrial product design,virtual reality,etc.In these areas,the retrieval of 3D models is becoming increasingly important.Researchers have proposed many 3D model retrieval algorithms.Among them,the 3D model retrieval algorithm based on multi-view has achieved superior performance,but the ability to mining of multi-view potential associations needs to be improved.At the same time,with the development of deep learning,some researchers have also applied it to 3D model retrieval,but currently there is a lack of comprehensive evaluation of the performance of deep learning in 3D model retrieval.Based on the above problems,this paper first comprehensively evaluates the performance of deep learning algorithms in view-based 3D model retrieval.In order to further improve the retrieval performance and mine multi-view potential association information,this paper designs non-local graph neural network and pairwise view weighted graph neural network,the three tasks of this paper are as follows:(1)Exploring the Deep Learning for View-based 3D Model Retrieval.Although deep learning has achieved great results in the areas of speech recognition and image recognition,and some researchers have applied it to 3D model retrieval.However,there is no comprehensive evaluation of the performance of deep learning in 3D model retrieval.Therefore,this paper systematically evaluates the performance of deep learning features and hand-crafted features in the following ways: 1)Comparison of different deep learning features with hand-crafted features;2)Evaluation of robustness of deep learning features;3)Evaluation of multi-view deep learning features.The results of large-scale experiments show that the performance of deep learning features is significantly better than the performance of hand-crafted features in 3D model retrieval,and deep learning features are more robust.In addition,multi-view deep learning features have lower computational complexity,better performance and better practicability.(2)Non-local Graph Neural Network for View-based 3D Model Retrieval.Although researchers currently use deep learning algorithms to extract features from multiple views,they have not considered the potential association between multi-view or simply cascading multi-view features.In order to make full use of the spatial relationship between different views,this paper proposes a 3D model retrieval method based on non-local graph neural network(NGNN).This method uses the Res Net as the basic network and embeds a non-local graph neural network into the basic network to learn the potential associations between multiple views.Finally,it obtains a highly responsive 3D shape descriptor through max-pooling.Experimental results on public datasets show that non-local graph neural network have very superior performance for mining multi-view potential associations.(3)Pairwise View Weighted Graph Network for View-based 3D Model Retrieval.Different views contain different information.Mining highly discrimination views and reducing the impact of confusing views is the key to improving retrieval performance.This paper designs a pairwise view weighted graph neural network(PVWGN)which contains three modules: original feature capture,multi-view relationship mining,and multi-view feature aggregation.Original feature capture is used to capture features of different levels.Non-local modules are used for multi-view relationship mining.The weighted view aggregation layer is used for different views weighting and view feature aggregation.Finally,the feature discrimination is further improved through the pairwise discrimination loss.Experiments on the public datasets show that PVWGN has achieved state-of-art performance in 3D model retrieval tasks.