A Graph Learning Model Based on Reconstruction Graph

Abstract

A graph learning model based on reconstructed graphs, which belongs to the field of image labeling, includes the following steps: by improving the nearest neighbor algorithm, finding the semantic nearest neighbor of the test image, constructing a similarity matrix for it, and clustering the images through a random dot product graph Classes, mine their internal connections to obtain a weighted similarity matrix, and then use graph learning algorithms to obtain preliminary image annotation results. Using the relationship between tags to label, considering the imbalance of co-occurrence between tags in this process, introducing the latest graph theory model, effectively solving the problem of tag imbalance. The random dot product map is used to reconstruct the transfer matrix of labels to solve the asymmetry problem of image label coexistence. Furthermore, a Naive Bayes nearest neighbor classifier is used to build a joint likelihood function between images and labels. Aiming at the feature of unbalanced classification of image tags, the present invention proposes an image tagging model based on a reconstructed graph model, which can effectively improve the recall rate of tags.

Description

technical field [0001] The invention belongs to the technical field of image labeling, and relates to a graph learning model based on reconstructed graphs, which is used to deal with the problem of image labeling under the background of big data. Background technique [0002] The advent of the big data era has brought many opportunities, but also brought more challenges. With the popularization of smart devices and the development of mobile networks, people are more and more fond of sharing images on the Internet, which produces a large number of images, which contain huge wealth. Therefore, efficient image analysis has become of great research significance hotspots. However, there are a large number of unlabeled images in these massive images, which poses a huge challenge to effectively mine the value of images. The graph learning model is a typical image annotation method, which is based on unsupervised, semi-supervised, and supervised learning strategies, and shares lab...

AI Technical Summary

Problems solved by technology

[0006] (2) Label co-occurrence imbalance phenomenon: In the image training set, there will be a phenomenon of simultaneous occurrence between labels

[0008] (1) The problem of low label recall rate: This problem refers to that when the image is labeled, the number of labels for some labels itself is too small, which will greatly reduce the probability of being selected during automatic labeling

[0009] (2) Label co-occurrence imbalance problem: This problem refers to the probability of co-occurrence between labels, but the co-occurrence probability between them is not exactly the same

[0010] (3) Visual ambiguity problem: This problem refers to the large difference between images with the same label, resulting in too large visual distance between them, which reduces the possibility of the image being selected as a semantic neighbor and affects the accuracy of image annotation

Images