Hyperspectral intelligent classification method based on prototype learning mechanism and multi-dimensional residual network

Abstract

The invention belongs to the technical field of image processing, and discloses a hyperspectral intelligent classification method based on a prototype learning mechanism and a multi-dimensional residual network. The method comprises the following steps: firstly, constructing a multi-dimensional residual network suitable for hyperspectral image features for extracting spectral and spatial featuresof a hyperspectral image; secondly, constructing a category prediction function based on a prototype learning mechanism, and replacing a softmax classifier used for traditional deep learning; and thenconstructing a novel prototype distance loss function, replacing a traditional softmax cross entropy loss function, and completing optimization and updating of multi-dimensional residual network parameters. The multi-dimensional residual network is introduced, a traditional softmax classifier and a softmax cross entropy loss function are abandoned, so that the complexity of the softmax cross entropy loss function is reduced. And a category prediction function and a prototype distance loss function based on a prototype learning mechanism are constructed and applied, so that the method has theadvantages of high precision for the hyperspectral image classification problem, high convergence rate in the training process, high robustness of the classification model obtained by training and thelike.

Description

technical field [0001] The invention belongs to the technical field of image processing, and further relates to a hyperspectral intelligent classification method based on a prototype learning mechanism and a multidimensional residual network in the technical field of hyperspectral image classification. The invention can be used to classify and identify different substances in hyperspectral images, and can play an important role in geological exploration, crop growth, camouflage revealing and the like. Background technique [0002] Hyperspectral imaging technology is a new technology gradually developed in the 1980s. It is a multi-dimensional information acquisition technology that combines traditional two-dimensional imaging technology and one-dimensional spectral detection technology. Hyperspectral images have extremely high spectral resolution and information richness, and can reflect the diagnostic spectral characteristics of target objects in more detail and accurately. ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to address the shortcomings of the above-mentioned existing deep learning hyperspectral classification technology that relies on softmax classifiers and softmax cross-entropy loss functions: the features learned by the classification model are not highly scalable, and the intra-class differences may be greater than the inter-class differences , the depth model is not robust, the classification accuracy is low, and the convergence speed of the training process is slow. A hyperspectral intelligent classification method based on a prototype learning mechanism and a multidimensional residual network is proposed.

Images