Hyperspectral Image Classification Method Based on Adaptive Multiscale Feature Extraction Model

Abstract

The invention discloses a hyperspectral image classification method based on an adaptive multi-scale feature extraction model, establishes a framework including two parts of a scale reference network and a feature extraction network, and introduces a conditional gating mechanism in the scale reference network, through three groups of modules Judgment, input the features into the corresponding scale extraction network, dig deep into the rich information contained in the hyperspectral remote sensing image, effectively combine the features of different scales, improve the classification effect, and generate a fine classification result map; in the feature extraction network, design a large-scale feature extraction network And the small-scale feature extraction network extracts feature information from two scales, comprehensively considers the heterogeneity of the dataset and the scale difference of the identified features, and can adaptively change the network structure to achieve multi-scale feature collaborative learning. The present invention simultaneously retains smaller-scale detail information and larger-scale spatial information of hyperspectral remote sensing images when performing multi-scale feature extraction, and meets the needs of hyperspectral remote sensing image target recognition and classification.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to an image classification method, in particular to a hyperspectral image classification method based on an adaptive multi-scale feature extraction model. Background technique [0002] Hyperspectral remote sensing is a multi-dimensional information acquisition technology that combines imaging technology and spectral technology. Since its development, it has shown great potential in all aspects and has become one of the most widely used fields of remote sensing. Hyperspectral remote sensing images have the characteristics of map-spectrum integration, high spectral resolution, and many bands. It can obtain almost continuous spectral characteristic curves of ground objects, and can select or extract specific bands to highlight object features as needed. Hyperspectral image data contains rich radiation, space and spectrum information, and is a comprehensive carrier of various in...

AI Technical Summary

Problems solved by technology

Existing methods (such as literature [2]) usually do not consider the differences in the scales of objects in the image, making it difficult to extract features of different scales at the same time, and large-scale features (such as grassland, ocean, etc.) can be better extracted, but Small-scale features (such as building edges, bricks and tiles in flower beds, etc.) are often ignored

Literature [3] proposed a two-stream learning model using convolutional neural network. However, due to the lack of multi-scale information, the classification effect is still limited.

Images