High-spectrum image classification method based on combined loss enhanced network

Abstract

The invention discloses a high-spectrum image classification method based on a combined loss enhanced network, and the method is characterized in that the method comprises the following steps: 1), PCA dimension reduction; 2), spatial domain block extraction; 3), coding path feature extraction; 4), classification task training target building; 5), decoding path feature extraction and reconstruction; 6), network combined training; 7), high-spectral testing classification. The method can achieve the combination of learning reconstruction loss and classification discrimination loss functions in the same network structure in a mode of end-to-end training, so the spatial spectrum information of a high-spectral image is used efficiently, and the learning of CNN for unimportant feature variables is automatically weakened, so as to reduce the complexity of a high-spectral classification model. Meanwhile, the dependence of a high-spectrum image classification method on a label sample is reduced, and the classification precision is improved.

Description

technical field [0001] The invention relates to the technical field of intelligent image processing, in particular to a hyperspectral image classification method based on a joint loss enhancement network. Background technique [0002] Hyperspectral image (HSI for short) has the unique advantages of high spectral resolution and map-spectrum integration, and has great application prospects in the fields of target tracking, environmental protection, agricultural monitoring, and weather forecasting. Classifying each pixel in a hyperspectral image is an important part of hyperspectral remote sensing applications and has great research significance. [0003] Traditional hyperspectral image classification methods often only use artificially extracted features in low-dimensional space. Typical methods mainly include: K-means clustering (K-means) method, logistic regression (Logistic regression, LR), support vector machine ( Support vector machine, SVM), etc., however, these classif...

AI Technical Summary

Problems solved by technology

[0003] Traditional hyperspectral image classification methods often only use artificially extracted features in low-dimensional space. Typical methods mainly include: K-means clustering (K-means) method, logistic regression (Logistic regression, LR), support vector machine ( Support vector machine, SVM), etc., however, these classification methods rely on shallow models, the invariance and discriminability of the extracted hyperspectral image features are poor, and it is easy to cause the curse of dimensionality (Hughes phenomenon), that is, with the hyperspectral image As the number of bands increases, the requirement for the number of training samples will increase sharply

Images