Natural image classification method and device based on multimodal matrix filling

Abstract

The present invention relates to an image classification method and device based on multimodal matrix filling. The method includes: performing feature extraction on labeled, unlabeled and tested natural image data to obtain different feature representations; using a matrix filling algorithm to generate The estimated label of each feature of the labeled data; the estimated label is linearly combined to approximate its corresponding known real label, and the combination coefficient is obtained; for various features, the matrix filling algorithm is used to predict the unidentified The label of the label and the label of the tested natural image data; using the combination coefficient to combine the labels of all the features predicted to obtain the label of the fusion of multiple features; classifying the natural image data based on the label of the fusion of multiple features. The invention is easy to realize, can obtain higher classification accuracy rate, and at the same time inherits the advantages of image classification based on matrix filling, and is suitable for the fields of network picture summarization and classification, image retrieval and the like.

Description

technical field [0001] The invention belongs to the technical field of image classification and multi-modal data analysis (multi-feature fusion), relates to a multi-label classification technology based on matrix filling, and specifically relates to an image classification method and device using multi-modal matrix filling. Background technique [0002] Different from images with single content and consistent form, such as faces and fingerprints, a natural image usually contains multiple objects, each of which presents different shapes. In natural image classification, it is often necessary to assign multiple class labels to an image. Such as figure 1 As shown, (a) "person" is riding a "bicycle", (b) "sky" and "ocean" often appear together, and (c) "dog" is a kind of "animal". Most of the traditional single-label classification (one sample has only one category label) algorithms cannot be directly used for multi-label classification. More feasible is the "one-to-many" str...

AI Technical Summary

Problems solved by technology

This approach will not only greatly reduce the calculation efficiency, but also lead to the problem of dimension explosion. At the same time, it lacks physical explanation and affects the classification accuracy.

Images