Semi-supervised image classification method and system for image embedded with LBP features

Abstract

The invention discloses a semi-supervised image classification method and system for an image embedded with local binary pattern (LBP) features. The method comprises the following steps: firstly, establishing an image library containing labeled samples and unlabeled samples; then constructing a convolutional neural network model, and training an initial model by using labeled samples in the image library; inputting the labeled samples and the unlabeled samples into the initial model, extracting feature vectors of the samples, and constructing an adjacent matrix Wcnn according to the feature vectors; then using the LBP features of the input sample image to construct an adjacent matrix Wlbp; adding the Wcnn and the Wlbp to obtain a new adjacent matrix W, constructing a graph according to the W, and obtaining a pseudo label of a label-free sample through label propagation; and finally, training a final model based on the initial model by using all samples and labels thereof in the image library, and carrying out image classification. According to the method, the image is constructed by introducing the LBP features of the image, so that the confidence of the label obtained through label propagation is higher, and the accuracy of image classification is improved.

Description

technical field [0001] The invention relates to a semi-supervised image classification method for a graph embedded with local binary pattern (LBP) features, which belongs to the field of image processing and pattern recognition. Background technique [0002] In the era of big data, the data obtained from the real world are usually huge in quantity and complex in structure, and most of them are unlabeled. The "label" here refers to the model output corresponding to the sample, which is the category of the sample in the classification problem. The traditional machine learning method is mainly supervised learning (supervised learning), which uses labeled (labeled) samples to train classifiers. In practical problems, in most cases, only a small part of the data is labeled. For example, in the image classification problem, in addition to the existing labeled data, new but unlabeled data will always be generated on the network. In order to take advantage of large amounts of unla...

AI Technical Summary

Problems solved by technology

Obviously, if only a small number of labeled samples are used, the model trained with it tends to have poor generalization performance; on the other hand, discarding a large number of unlabeled samples is a great waste of resources

Images