Multidirectional information propagation convolutional neural network construction method for hyperspectral image classification

Abstract

The invention discloses a multidirectional information propagation convolutional neural network construction method for hyperspectral image classification. The method comprises the steps that an inputend is a local three-dimensional hyperspectral data cube sample taking a target pixel as a center; a deep neural network is composed of a two-dimensional convolution multilayer perceptron between hidden layer units, a two-dimensional convolution perceptron in the hidden layer units, a pooling layer and a full connection layer; the two-dimensional convolution perceptron in the hidden layer units is used for fragmenting a hidden layer internal feature map in the row or column direction and executing the fragment-by-fragment convolution among the feature fragments in the upper, lower, left and right directions so as to transmit the space information of pixels in different directions; and an output layer is a category probability vector of the input spectral pixels. The network is different from a classic convolutional network, a spatial information propagation mechanism between feature channels is formed in a hidden layer, spatial spectrum features with higher discrimination can be learned, the network is applied to hyperspectral supervised classification, and the supervised classification capability under a small number of samples is greatly improved.

Description

technical field [0001] The invention relates to deep neural network technology, in particular to a multi-directional information propagation convolutional neural network construction method for hyperspectral image classification. Background technique [0002] In recent years, convolutional neural network, as a popular deep learning framework, has gradually become a powerful tool in hyperspectral image analysis, and its application prospects in the field of hyperspectral classification are very promising. Compared with methods based on shallow representation learning, convolutional neural networks implemented by deep convolutional perceptrons can adaptively learn hierarchical representations from low-level features to high-level features, and then sequentially identify the most discriminative features for high-level Supervised classification task of spectral images. For hyperspectral image classification tasks, the challenges lie in the following aspects. First, pixels are ...

AI Technical Summary

Problems solved by technology

Although these methods are effective in improving hyperspectral classification, there are still many deep features in hyperspectral images that have not been utilized, especially in terms of spatial features, which are far from enough.

Images