Hyperspectral image noise label detection method based on super-pixel weight density

Abstract

The invention discloses a hyperspectral image noise label detection method based on super-pixel weight density. The method comprises the following steps: firstly, introducing an entropy rate superpixel (ERS) algorithm into a step of obtaining context data information for training samples (with noise tags); secondly, obtaining Euclidean distance measurement through a super-pixel region; next, introducing a Gaussian weighting algorithm into Euclidean distance calculation, defining a weighting coefficient of distance information of each class, finally, adaptively detecting a training sample through a clustering algorithm based on a density peak value, and evaluating the effectiveness of the proposed detection method by utilizing a support vector machine. Experiments carried out on several real hyperspectral data sets show that the provided label detection method can effectively improve the performance of a classifier with label noise pollution in an initial training sample, and has very great theoretical significance and practical application value.

Description

technical field [0001] The invention relates to a hyperspectral image noise detection method, in particular to a hyperspectral image noise label detection method based on superpixel weight density. Background technique [0002] Hyperspectral images have complex spectral bands and spatial information features, and have great application value in scene recognition and other aspects. However, the acquisition of images is often affected by factors such as atmospheric interference and manual labeling errors, making the existing hyperspectral image training samples contain a large number of noise labels. Therefore, effectively detecting and removing noisy labels of training samples has become a research hotspot in the field of hyperspectral image supervised classification in recent years. In the supervised classification with noisy labels, based on the superpixel segmentation method, the local similarity space and spectral information can be fully used to improve the classificati...

AI Technical Summary

Problems solved by technology

[0003] Aiming at the problem that training samples containing label noise in supervised classification weaken the classification performance of hyperspectral classifiers, the present invention provides a hyperspectral image noise label detection method based on superpixel weight density

The invention integrates the superpixel segmentation algorithm, the Gaussian weighting rule, and the density peak clustering algorithm, thereby strengthening the measurement information of similar pixel similarity, and can effectively detect and remove noise pixels, and solves the problem of high The problem of low classification performance of the classifier when the spectral image training samples contain many noisy labels

Images