VQ-VAE-based frequency height graph similarity retrieval system and method

Abstract

The invention belongs to the technical field of ionospheric frequency height maps, and particularly relates to a VQ-VAE-based frequency height map similarity retrieval system and method, through a codebook optimization method combining a Gaussian mixture model and orthogonal constraints, frequency height maps are clustered by using GMM, and the orthogonal constraints are introduced in the clustering process, so that the expression ability and feature capture precision of a codebook are improved. The introduction of the orthogonal constraint term is helpful for reducing overlapping among clusters, and the capability of distinguishing complex image features by the codebook is improved. According to the method, high-efficiency image similarity calculation is achieved by calculating the Euclidean distance between low-dimensional vectors generated by VQ-VAE, and in GMM and orthogonal constraint optimization, convergence and stability of the model are ensured by adjusting a regularization coefficient (lambda) and the maximum number of iterations. The mechanism can further improve the robustness of the model, and has strong generalization ability in practical application.

Images