Cross-modal hash retrieval method based on supervision graph embedding

Abstract

The invention discloses a cross-modal hash retrieval method based on supervision graph embedding, and belongs to the technical field of multimedia retrieval. Constructing an image and text modal data set, and dividing the image and text modal data set into a training set and a test set; extracting features of all image and text samples of the training set and the test set, and mapping the extracted features to a nonlinear kernel space by using a radial basis kernel function to improve the characterization capability of the features; generating a pairwise similarity matrix of the samples by utilizing the category labels, and further generating a Laplacian matrix; generating a semantic subspace by utilizing the category label; learning a mapping matrix for image and text modalities by utilizing intra-modal similarity keeping based on an image embedding method and inter-modal similarity keeping based on a semantic subspace; learning an orthogonal rotation matrix to minimize a quantization error; and an efficient iterative discrete optimization algorithm is utilized to reduce the calculation complexity of the training process.

Description

technical field [0001] The invention relates to a cross-modal hash retrieval method based on supervision graph embedding, belonging to the technical field of multimedia retrieval. Background technique [0002] In recent years, graph embedding based hashing methods have attracted much attention due to their effectiveness in cross-modal retrieval. The hash method is to map similar samples in the original space into similar hash codes, and then calculate the Hamming distance between the hash codes by XOR operation, and use the Hamming distance to measure the similarity between samples, which can significantly reduce the Computational complexity and memory overhead. Inspired by this, researchers have proposed many hashing methods for large-scale retrieval tasks in recent years. However, most methods are only applied to unimodal data, i.e. the modality type of the retrieved sample is the same as the modality type of the query sample, e.g. using images to retrieve images. But t...

AI Technical Summary

Problems solved by technology

However, the threshold operation will bring quantization error, which usually leads to the reduction of the effectiveness of the hash code; 3) most hashing methods based on graph embeddings fail to fully utilize the class labels of the samples during the training process, resulting in their retrieval results hard to satisfy

Images