Hyperspectral image classification method based on novel activation function

Abstract

The invention provides a hyperspectral image classification method based on a novel activation function, and the method comprises the steps: carrying out the dimensionality reduction and whitening operation of a hyperspectral data set in each layer of a random block network, obtaining a feature matrix, obtaining a feature map through an s-ReLu correction linear unit activation function, and finally carrying out the ground object classification of a normalized feature fusion image, and obtaining a classification result graph. According to the method, the linear unit of the ReLu activation function is optimized by using the softsign activation function, and a new s-ReLu correction linear unit activation function is obtained, so that the value domain is smoother and more flexible when being 0-0.5, and the problem that other activation functions cannot better provide appropriate feature activation for pixel points is solved, so that the feature expression ability is improved, and the feature extraction efficiency is improved. And the classification precision of the classification model is improved.

Description

technical field [0001] The invention relates to the fields of image processing and computer vision research, in particular to a hyperspectral image classification method based on a novel activation function. Background technique [0002] In recent years, Hyperspectral Image (HSI) imaging technology has made breakthroughs in many fields, and has a wide range of applications in agriculture, mineralogy, environmental monitoring, and astronomy. In practical applications, the requirements for classification accuracy of ground objects are getting higher and higher. For example, precision agriculture requires precise classification of crop species for precise fertilization and accurate yield estimation; another example is precision forestry, which requires high-precision classification of forest areas by tree species in order to achieve precise detection, precise management, and precise breeding. Therefore, how to improve the classification accuracy of the HIS model has become a k...

AI Technical Summary

Problems solved by technology

[0004] With the continuous development of machine learning and deep learning for a long time, the commonly used activation functions can be divided into the following three types. The first one is the sigmoid and tanh activation functions, both of which are S-type saturation functions. When the input tends to be positive or negative When it is infinite, the phenomenon of gradient dispersion is prone to occur; the second is the softsign activation function, softsign is an improved version of the tanh function, the value range is (-1,1), centered on zero, the function curve is more gentle, which can ease the gradient The problem of disappearance; the third is the ReLu activation function, the ReLu activation function is more concise and the convergence speed is faster, and the derivative is always 1 without gradient dispersion, but neurons with large gradients may occur when neurons pass by "Necrotic" situation

Images