Robust self-adaptive semi-supervised image classification method and device, equipment and medium

Abstract

Embodiments of the invention disclose a robust self-adaptive semi-supervised image classification method and device, equipment and a medium. The method comprises the following steps of: integrating robust self-adaptive embedded label propagation and self-adaptive weight construction into a uniform semi-supervised learning framework, and minimizing an embedded characteristic and embedded label-based reconstruction error; changing an original predicted label set into a preset label space by utilizing robust projection, so as to classify each label in the original predicted label set; decomposingan original data set into a denoising result expression item and a noise fitting error item, and carrying out self-adaptive weight construction and self-adaptive label propagation on the denoising result expression item; and integrating a regressed label approximate error item into the semi-supervised leaning framework, and carrying out combined optimal learning to obtain a projection classifiermatrix. According to the method, the image classification ability and pattern classification prediction ability are effectively enhanced, and benefit is brought to enhance the image classification correctness.

Description

technical field [0001] Embodiments of the present invention relate to the technical fields of data mining and image recognition, in particular to a robust adaptive semi-supervised image classification method, device, equipment and computer storage medium. Background technique [0002] In the field of data mining and image recognition, image-based semi-supervised learning has received great attention because it can use a small amount of labeled data and a large amount of unlabeled data for data representation and classification. In recent years, the image semi-supervised learning model for label estimation, that is, the semi-supervised classification method based on the label propagation model, has attracted the attention and research of many researchers because of its high efficiency. [0003] Existing label propagation methods include GFHF (Gaussian Fields and Harmonic Function, Gaussian Fields and Harmonic Function), LNP (Linear Neighborhood Propagation, linear field propa...

AI Technical Summary

Problems solved by technology

Most of the current methods divide the weight construction process and the label propagation process into two independent steps, but this operation cannot ensure that the weight coefficients obtained in advance are optimal for subsequent label estimation, and may reduce its learning performance; each The neighborhood of points is usually determined using K-neighborhood or ε-neighborhood, however, due to the complex distribution of real data, estimating an optimal K or ε value is challenging in reality; most existing label propagation models Pre-settlement of weight coefficients based on raw data that often contains redundant information, unfavorable features, and noise, which can also potentially degrade learning performance; existing methods often have unfavorable mixed signals in the predicted labels due to lack of proper constraints , however the unfavorable signals contained in it may directly lead to inaccurate label prediction and classification results

Images