Cosine measurement supervised deep hash algorithm with balanced similarity

Abstract

The invention discloses a cosine measurement supervised deep hash algorithm with balanced similarity, and belongs to the field of image retrieval. The deep supervised hash has the advantages of low storage cost, high calculation efficiency and the like. However, similarity preserving, quantization error, and imbalance data are still huge challenges in deep supervised hashing. The invention provides a deep hash scheme for pairwise similarity preservation, and solves the problem. According to the method, a deep network is used as a basic model to extract features, and a hash layer is used to replace a final classification layer to enable the final classification layer to output hash codes. According to the method, a loss function is designed, semantic similarity can be effectively kept in the training process, and the problems of class imbalance, difficulty and quantitative loss are solved. When the hash code obtained by the method is used for image retrieval, the retrieval accuracy canbe effectively improved for an extremely unbalanced data set.

Description

technical field [0001] The invention relates to the field of image retrieval, and in particular relates to a cosine metric supervised deep hash algorithm with balanced similarity. Background technique [0002] With the rapid development of multimedia processing technology, large-scale image search has been widely used in our daily life. As one of the most effective technologies, hashing technology has attracted more and more attention from academia and industry due to its storage and computing advantages. It aims to map high-dimensional images to compact binary codes while preserving image correlation. In particular, deeply supervised hashing may improve retrieval performance by integrating image features and hash coding through end-to-end learning. [0003] Hash embedding is generally implemented with discrete optimization, which is a standard NP-hard problem. Deeply supervised hashing methods mostly employ continuous relaxation, replacing discrete encoding with approxim...

AI Technical Summary

Problems solved by technology

Although an easy pair contributes less to the global gradient than a hard pair, the total contribution of a large number of easy pairs is cumulative and may exceed that of a small number of hard pairs

This phenomenon makes the model converge slowly, may fall into local optimum, and even reduce the retrieval performance to a certain extent

Images