Abnormity detection method and device on basis of non-negative matrix factorization

Abstract

The invention discloses an abnormity detection method on the basis of non-negative matrix factorization, which comprises the following steps: preprocessing read hyperspectral images to obtain hyperspectral images which are subjected to noise removal; carrying out vector conversion on the obtained hyperspectral images which are subjected to noise removal so as to obtain a two-dimensional initialization matrix V; then carrying out linear decomposition on the two-dimensional initialization matrix V to generate a random initialization basis matrix W and a coefficient matrix H; according to a non-negative matrix factorization multiplicative algorithm, carrying out iteration on the random initialization basis matrix W and the coefficient matrix H to obtain hyperspectral images with a plurality of wave bands; finally, according to a local self-adaptive kernel density estimation operator, processing the hyperspectral images of which the wave band has the greatest quantity of abnormal information in the hyperspectral images with a plurality of wave bands so as to obtain images of which abnormal targets are detected. The invention also discloses an abnormity detection device on the basis of non-negative matrix factorization. By the abnormity detection method and the abnormity detection device on the basis of non-negative matrix factorization, a great quantity of redundant wave bands and noise information can be eliminated so as to effectively improve efficiency of abnormity detection.

Description

technical field [0001] The invention belongs to the technical field of remote sensing data processing, and in particular relates to an anomaly detection method based on non-negative matrix decomposition and a device thereof. Background technique [0002] Remote sensing technology is one of the major technological breakthroughs that humans have made in earth observation at the end of the 20th century, and hyperspectral remote sensing technology is the frontier technology of current remote sensing. A hyperspectral image is a three-dimensional image that has a spectral dimension in addition to the two-dimensional plane on traditional images. Each band of a hyperspectral image can be regarded as an image independently, and the images of all bands are superimposed to form an image cube. Theoretically, it has been proved that the data of each band between the spectral dimensions in the hyperspectral data cube has a strong correlation, and there is almost no data in the high-dimen...

AI Technical Summary

Problems solved by technology

This method is much more efficient than the PCA algorithm, but it requires a complex preprocessing process, and the algorithm is highly random, and the results obtained have a lot to do with the initial value, so it is not easy to implement

Images