Hyperspectral image classification method based on Transform-CNN-KAN collaborative framework

Abstract

The invention discloses a hyperspectral image classification method based on a Transform-CNN-KAN (convolutional neural network-KAN) collaborative framework, and the method comprises the steps: carrying out the feature extraction of a single pixel through two channels of CNN and KAN, can maintain the discriminative features of the single pixel while reducing the redundant information, facilitates the subsequent simplification of a differential attention mechanism for global feature aggregation, and improves the classification precision of the hyperspectral image. A single-layer single-head simplified differential Transform is used for aggregating global spectral information, the complexity of the Transform can be reduced through the simplification thought, the problem of ineffective attention diffusion can be solved through the differential thought, more attention is put on similar ground features, and therefore intra-class differences are reduced, and inter-class differences are expanded; according to the method, multi-scale space information is aggregated through two-channel different-scale 2D depth separable convolution, the calculation amount of the used depth separable convolution is smaller than that of common 2D convolution, and the calculation efficiency is higher.

Images