Discrete supervised cross-media Hash retrieval method based on collaborative matrix decomposition

Abstract

The invention relates to a discrete supervised cross-media Hash retrieval method based on collaborative matrix decomposition, which comprises the following steps of: 1) establishing a cross-media retrieval database, and dividing sample pairs into a training set and a test set; 2) extracting features of all sample pairs in the training set and the test set, and performing mean value removal; 3) projecting features and category labels of the samples to a low-dimensional feature space and a Hash code by using collaborative matrix decomposition and semantic embedding respectively, learning an orthogonal rotation matrix to construct semantic association of the low-dimensional features and the Hash code, and learning a Hash function for each mode; 5) generating a hash code of the test sample byutilizing the learned hash function; 6) taking the samples in the training set as to-be-retrieved samples, taking the samples in the test set as query samples, and calculating the Hamming distance between the query samples and the to-be-retrieved samples; 7) arranging according to a Hamming distance descending order, and returning the first r heterogeneous samples as retrieval results. According to the cross-media retrieval method based on the Hamming distance, cross-media retrieval can be achieved, occupied resources in the training process are few, the accuracy rate is high, and the cross-media retrieval method based on the Hamming distance has wide application prospects.

Description

technical field [0001] The invention relates to the fields of multimedia retrieval and artificial intelligence, in particular to a discrete supervised cross-media hash retrieval method based on collaborative matrix decomposition. Background technique [0002] With the rapid growth of data volume on the web, how to retrieve semantically similar samples in large-scale data becomes a challenge. On the one hand, due to the high time complexity and storage overhead, it is difficult to directly apply the traditional nearest neighbor retrieval method to large-scale data. On the other hand, the media types of samples on the Internet are diverse, and how to measure the similarity between heterogeneous samples across the gap between media types has become a challenge. Cross-media hashing maps the high-dimensional features of heterogeneous samples to a shared low-dimensional Hamming space to measure the similarity of heterogeneous samples. Due to its high efficiency and effectiveness,...

AI Technical Summary

Problems solved by technology

However, there are two problems in this type of method: 1) Using labels to construct the similarity matrix between two samples will cause the loss of category information; 2) The size of the similarity matrix between two samples is , however in large-scale applications, The value of is very large, so it will bring excessive memory overhead and computational complexity, making it lose the ability to expand

Most existing methods only keep learning hash codes based on the semantic similarity of class labels. However, due to the complex semantic association between heterogeneous samples, it is difficult for many heterogeneous samples with the same class labels to be mapped into similar Hamming codes.

In addition, due to the discrete constraints of the hash code, the objective function of the hash method is difficult to solve directly

Images