Shoeprint image clustering method guided by interactive text semantic attributes

Abstract

The invention provides a shoeprint image clustering method guided by interactive text semantic attributes. The method mainly comprises the steps of shoeprint image clustering based on content and shoeprint image clustering based on interactive text semantic attributes. Shoeprint image clustering based on interactive text semantic attributes mainly comprises the steps of attribute definition, automatic sample selection of an attribute classifier, labeling of the semantic attributes of the samples, online training of the attribute classifier, category refinement based on attribute classification, sample increment selection and re-clustering and the like. According to the method, the interactive text semantic attributes are introduced for guiding, so that the influence of parameters on a clustering result is improved, the subjective intention of a user is reflected interactively, and a result which better conforms to subjective evaluation of people can be obtained.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a shoe print image clustering method guided by interactive text semantic attributes. Background technique [0002] For the sole pattern images with high similarity, it is often difficult to cluster. The shoe print dataset has a large number of categories, so it is not suitable for existing semi-supervised clustering algorithms. And the shoe print data set has the problem of few samples of each type and uneven distribution of samples, so it is not suitable for clustering algorithms based on deep learning. [0003] The existing shoeprint clustering algorithms are often based on image content, and the result of clustering only based on image content has a semantic gap to a certain extent. In addition, the current traditional clustering algorithm and the clustering algorithm based on deep learning both need to set a large number of parameters, and the parameter adj...

AI Technical Summary

Problems solved by technology

The shoe print dataset has a large number of categories, so it is not suitable for the existing semi-supervised clustering algorithm

And the shoe print data set has the problem of few samples of each class and unbalanced sample distribution, which is not suitable for clustering algorithms based on deep learning

[0003] Existing shoeprint clustering algorithms are often based on image content, and the result of clustering only based on image content has a semantic gap problem to a certain extent

In addition, the current traditional clustering algorithm and the clustering algorithm based on deep learning both need to set a large number of parameters, and the parameter adjustment is complicated. Different parameter combinations have a great impact on the quality of the clustering results, which may result in high purity and low recall rate. Or the phenomenon of high recall and low purity

Images