An image retrieval method based on hyperbolic curvature semantic alignment hashing

Abstract

This invention discloses an image retrieval method based on hypercurvature semantic alignment hashing, belonging to the field of computer vision technology. The key technical points of this method are: it includes five steps: data preparation and preprocessing, constructing a feature extraction backbone network, executing a hypercurvature collaborative feature extraction strategy, hash code generation, and establishing a loss function. By constructing a hypercurvature collaborative feature extraction network to adaptively adapt to the Euclidean and quasi-hyperbolic geometric properties of visual data, and combining semantic alignment ranking and polarization quantization joint loss, the method effectively solves the problems of feature geometric heterogeneity and training target misalignment, achieving end-to-end image retrieval with both binary compactness and high semantic discriminative power.

Images