面向图像检索的混合学习索引方法OA

In order to solve the problem of semantic gap and dimensional disaster in the processing of large-scale high-dimensional data by traditional image retrieval methods,we propose a Multimodal Dynamic Learned Index(MDLI).The method achieves a breakthrough through a three-level synergy mechanism.Firstly,the hierarchical adaptive weighting module is designed to achieve multi-scale feature fusion,and the local details and global semantics of different levels of ResNet are integrated.Secondly,the improved Graph Attention Network(GATv2)is introduced to dynamically model the complex relationship between images,and the top-20 edge sparsity strategy is combined to improve the computational efficiency.Finally,a hybrid index architecture is constructed,which organically combines the learned MLP index with the traditional VP-Tree,and optimizes the retrieval performance through the dynamic routing mechanism.Ex-periments on MNIST,CIFAR-10 and ImageNet-1K datasets show that the proposed method is significantly better than the existing methods in terms of retrieval accuracy and efficiency,and provides a solution for large-scale image retrieval that takes into account both accuracy and efficiency.