Super-pixel classification method based on semi-supervised K-SVD and multi-scale sparse representation

Abstract

The invention discloses a super-pixel classification method based on semi-supervised K-SVD and multi-scale sparse representation, and the method comprises the steps: firstly carrying out the semi-supervised K-SVD dictionary learning of a training sample of a hyperspectral image, and obtaining an over-complete dictionary; secondly, taking the training sample and the over-complete dictionary as input, and performing super-pixel multi-scale sparse solution to obtain a sparse representation coefficient matrix of the training sample; and finally, obtaining a super-pixel classification result through a residual error method and a super-pixel voting mechanism according to the obtained sparse representation coefficient matrix and the over-complete dictionary. The method has good capabilities of removing salt and pepper noise and enriching training samples. A very stable classification result can be achieved under the condition of various sample quantities. The method is of great significance in solving the problem of salt and pepper noise and the problem of high-dimensional small samples in the field of hyperspectral image classification and how to effectively utilize spatial information through a classification algorithm based on sparse representation.

Description

technical field [0001] The invention relates to the technical field of hyperspectral image classification, and more specifically relates to a super-pixel classification method based on semi-supervised K-SVD and multi-scale sparse representation. Background technique [0002] Over the years, remote sensing image classification has played a vital role in a variety of applications such as environmental damage assessment, crop growth regulation, land use monitoring, urban planning, and reconnaissance. Compared with single-band panchromatic images and multispectral images, hyperspectral images (HSI) can detect and distinguish objects with higher accuracy because of their higher spectral resolution. [0003] In a hyperspectral image, the spectral data of each pixel is a high-dimensional vector, and hundreds of data dimensions represent the spectral response of hundreds of bands. Hyperspectral image classification is mainly to classify each pixel according to the spectral informat...

AI Technical Summary

Problems solved by technology

However, in these two algorithms, there are still many problems in the selection of the adjacent space: on the one hand, the shape of the adjacent space is rectangular, and the objects in the rectangular window are unknown, and the algorithm only assumes that the objects in the rectangular window are uniform. Yes, when the scale is large, this assumption becomes very dangerous; on the other hand, the scale of the adjacent space area is single and needs to be set in advance. Different application environments have different optimal scales, so how to configure this scale is a problem. with many difficulties

Images