DESIGN AND RESEARCH OF A MULTI-VIEW GRAPH DEEP LEARNING 3D MODEL RETRIEVAL SYSTEM BASED ON FUSION VISION-TRANSFORMER

The development of computer vision has made three-dimensional models play a crucial role in the field of image processing. However, compared to 2D models, 3D models have more features, making it difficult to extract features and mine correlation information between features. Based on this, this study is based on a multi-perspective graph convolutional neural network, which uses an image entropy weight pooling layer to improve the original view pooling layer. It assigns a weight based on image entropy to each perspective image, and then performs view pooling operations. The Vision-Transformer module is embedded into a multi-perspective graph convolutional neural network to mine information associations between multi-view graphs. The results show that the multi- perspective graph convolutional neural network model fused with Vision-Transformer is more concentrated in classifying features of the same category in the view graph, and there is a significant distance difference between different features. The multi-perspective graph convolutional neural network model fused with Vision-Transformer achieves accuracy of 89.0%, 92.0%, 94.0%, and mean average precision values of 80.0%, 85.0%, and 88.0% when the number of view images is 6, 10, and 14. This study improves the retrieval accuracy of 3D models and has certain reference value in the field of computer vision.