Shape adaptive convolution deep neural network method for hyperspectral image classification

Abstract

The invention discloses a shape adaptive convolution deep neural network method for hyperspectral image classification. The method comprises the following steps: adopting a spatial structure information learning branch; using a shape self-adaptive convolution kernel based on a guide graph and the shape self-adaptive convolution kernel can be trained; a spectral dimension one-dimensional convolution layer and a spatial dimension two-dimensional convolution layer forming a space-spectrum feature extraction unit, and each unit having two inputs, namely a feature map and a guide map; wherein the deep network is formed by stacking a plurality of space-spectrum feature extraction units, and a skip layer connection is established between every two feature extraction units; wherein the network loss function is weighted cross entropy. through learning the spatial correlation between adjacent pixels in the space-spectrum data is learned, the receiving domain shape of convolution operation can beadaptively adjusted according to the spatial structure relationship between explicit definition pixels, the defect that anisotropic characteristics cannot be captured by fixed square convolution is overcome, and the method has excellent classification and generalization performance for hyperspectral images with different resolutions and different scene complexities.

Description

technical field [0001] The invention relates to hyperspectral image classification technology, in particular to a shape adaptive convolution deep neural network method for hyperspectral image classification. Background technique [0002] The hyperspectral camera can obtain cube-like "map-spectrum integration" data rich in material information. It can have nanometer (nm)-level spectral resolution in the visible light-near-infrared, short-wave infrared, and even mid-infrared and thermal infrared bands, with up to hundreds of A continuous, narrow-band spectral band image is widely used in military reconnaissance, environmental monitoring, geological exploration and target detection and other fields. Among them, hyperspectral image (HSI) supervised classification is one of the most important research contents in this field. [0003] In the past decade, researchers have proposed many supervised classification methods for HSI. From simple models based on statistics to complex me...

AI Technical Summary

Problems solved by technology

The obvious result of this CNN defect is that the classification map becomes too smooth and details of many scenes are lost, and the classification effect of HSI with rich scene details is poor.

Images