Decorrelation clustering method and device under data selection deviation

Abstract

The embodiment of the invention provides a decorrelation clustering method and device under data selection deviation, and the decorrelation clustering method comprises the steps: obtaining a pluralityof images with deviation, and enabling the images to serve as a sample set; based on the sample set, combining and optimizing a post-weighting clustering algorithm and a decorrelation regular term toobtain an optimal post-weighting clustering algorithm, obtaining the optimal post-weighting clustering algorithm by calculating the post-weighting clustering algorithm for multiple times, and obtaining the post-weighting clustering algorithm by using the weight of each sample obtained by learning the decorrelation regular term to weight the clustering algorithm, wherein the weight of each sampleis obtained by learning the weight of each sample of each image in the sample set by using a decorrelation regular term; and obtaining an optimal weighted clustering algorithm when a current clustering center and a cluster contained in the current weighted clustering algorithm are not the first clustering center and the cluster and the difference between the current clustering center and the cluster and the last clustering center and the cluster is smaller than a threshold value, so that the clustering center and the cluster, not affected by deviation, of an image are determined.

Description

technical field [0001] The invention relates to the technical field of image classification, in particular to a decorrelation clustering method and device under data selection bias. Background technique [0002] Usually in the process of image clustering, data selection bias is prevalent in real-world situations. And data selection bias may lead to spurious associations between image features. Assuming that a spurious feature is incorrectly identified as being associated with an important feature, the role of such meaningless features will be enhanced in image clustering due to the presence of spurious correlations, making the inherent data distribution impossible to obtain reveal. Examples are as follows: [0003] See for example figure 1 , when collecting image data, the foreground is the picture of the target object, such as a picture of a dog and a picture of a cat, but the background of the target object is not the same, for example, there are more pictures of dogs ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the embodiment of the present invention is to provide a decorrelation clustering method and device under data selection bias, to solve these biased problems in the prior art Clustering of data will lead to technical problems that the clustering effect will decline

Images