Image retrieval method utilizing deep semantic to rank hash codes

Abstract

The invention discloses an image retrieval method utilizing deep semantic to rank hash codes. The method includes the following steps that partial images in a multi-label image data seat serve as a training set, and the left images serve as an image testing set; a deep convolutional neural network is utilized to construct a deep hash function; a semantic similarity rank among the images is constructed according to multi-label information of the images; the deep hash function is optimized by using an agent ranking loss function based on a triple as an actual model objective function and using a stochastic gradient-descent; the hash codes of the images are calculated through the learned deep hash function, and the images can be retrieved by testing the Hamming distance between the hash codes in the images and the hash codes of each image in the training set through calculation. The method can reserve multi-stage similarities of the multi-label images in the semantic space and express and hash the codes in combination with depth features of the learned images, thereby avoiding loss of semantic information.

Description

technical field [0001] The invention relates to the field of pattern recognition and machine learning, in particular to image hash coding and retrieval. Background technique [0002] The explosive growth of network images makes content-based image retrieval more and more important. Hash coding has been widely used in large-scale image retrieval due to its high efficiency and easy storage. Nevertheless, most of the hash coding methods based on semantic tags only solve the problem of binary similarity measure (i.e. similar or dissimilar), but they cannot handle the multi-level semantic similarity measure in multi-label images well. . In addition, the process of most hash coding methods is generally to first extract the hand-designed feature expression, and then learn a deep hash function on this basis. These hand-designed features tend to describe appearance rather than semantic information, and are not suitable for tasks involving complex semantic structures. Contents of...

AI Technical Summary

Problems solved by technology

Meanwhile, an algorithm based on a surrogate ranking loss function is used to solve the multivariate non-smooth ranking metric optimization problem during learning

Images