Multi-label image annotation result fusion method based on rank minimization

Abstract

The invention relates to a multi-label image annotation result fusion method based on a rank minimization optimization algorithm. The method comprises the following steps: (1), extracting various feature representations of a training set image, wherein the training set image has a semantic annotation word which is given in advance; (2) in the various feature representations, training each supervised learning image annotation model; (3) for a new image without a semantic annotation word, using the same method for extracting various features of the image, and using the features for being respectively input to the corresponding supervised learning image annotation models to predict multi-label results; (4) utilizing the rank minimization algorithm for fusing the multi-label results output by the various models to obtain a more accurate annotation result. According to the multi-label image annotation result fusion method based on the rank minimization, the complementarity of the image annotation models under the various feature representations is fully utilized, the rank minimization algorithm is utilized for reducing the number of prediction mistakes in the fused annotation result, and therefore the final image annotation result is more accurate.

Description

technical field [0001] The invention relates to a fusion method of image labeling results, in particular to a fusion method of multi-label image labeling results based on rank minimization. Background technique [0002] With the popularization of digital cameras and social networking applications, people are more and more used to posting the images they take on the Internet. In order to manage and retrieve massive images on the Internet more conveniently, automatic image annotation is an effective tool. The basic task of an image tagging program is to model the relationship between the low-level visual features of an image and the high-level semantically tagged words. The supervised learning image tagging model uses images with semantic tagging words as training images, first extracts the underlying visual features of the training images, and then uses these feature representations and corresponding semantic tagging words as input to train a supervised learning image taggin...

AI Technical Summary

Problems solved by technology

Considering that the results of different models will be in different numerical ranges, the normalization method will be different for different models, increasing the complexity of decision-level fusion

Furthermore, the fusion effects of different fusion rules are not the same, so there will be doubts about which fusion rules are the best for which model fusion.

In addition, since a single discriminative model has its own prediction error, blindly using rules to fuse the results of multiple models will also introduce their own errors at the same time

Images