Category label recovery method for low-rank image feature analysis

Abstract

The invention relates to a category label recovery method for reliable image classification, and belongs to the technical field of machine vision. The invention discloses a category label recovery method for low-rank image feature analysis by combining a label propagation process, PCA dimension reduction and LDA classification. According to the method, a label propagation process is introduced toobtain a data set after label recovery, then PCA reliable dimensionality reduction is executed to reduce the dimensionality of a reliable data set, meanwhile, information with the largest contributionto variance in the reliable data set is kept, finally, LDA features are extracted, and low-rank features for discriminant analysis are obtained. According to the method, a data set containing missinglabels is preprocessed, the missing category labels are restored, and the most discriminative feature is extracted through reliable dimension reduction, so that a subsequent nearest neighbor classifier is more accurate and reliable; and the robustness of the sample label data is improved, so that the classification model is more effective.

Description

technical field [0001] The invention relates to a category label recovery method for low-rank image feature analysis, which belongs to the technical field of machine vision. Background technique [0002] Linear discriminant analysis (LDA) can use the known category labels of sample images to find the projection subspace that is most helpful for image classification, which belongs to a supervised machine learning method. LDA extracts the most discriminative low-rank features from high-dimensional data. These features help to gather all samples of the same category and separate samples of different categories as much as possible. The feature with the largest in-divergence ratio. Efficient preprocessing of image datasets is quite difficult. Generally, an effective method is to keep the main information contained in the dataset as much as possible and reduce its dimensions. Typically, principal component analysis (PCA) can perform effective dimensionality reduction analysis on...

AI Technical Summary

Problems solved by technology

[0003] Label information plays a vital role in LDA-based image classification, but the actual obtained datasets often have the problem of missing labels.

If this problem is ignored, the trained classification model may have various problems. If the sample images without labels are discarded, the model will be underfitted or the recognition rate will be low.

Images