Center-Guided Dynamic Convolutional Network for Hyperspectral Image Classification

In hyperspectral image (HSI) classification, effectively modeling the relationship between a target pixel and its surrounding context remains challenging. While Transformer-based approaches achieve promising results through self-attention mechanisms, their quadratic computational complexity limits practical deployment. In this paper, we propose a novel Center Guided Dynamic Convolutional Network (CGDC-Net) that enhances CNN architectures with input-adaptive convolutions and center-guided attention mechanisms. CGDC-Net introduces the Central Dynamic Convolution (CDConv) module that employs dynamically modulated kernels adapting to each input sample based on multi-scale contextual information centered around the target pixel. Through efficient tensor decomposition, CD Conv generates sample-specific kernel modulations while maintaining computational efficiency. Building upon CDConv, we further incorporate Center-Guided Convolutional Block Attention (CGCBA) that integrates center-guided channel attention (CGCA) and center-guided spatial attention (CGSA) to explicitly emphasize the target pixel while capturing discriminative spectral-spatial features. By combining dynamic convolutions with attention mechanisms, CGDC-Net achieves enhanced feature representation while maintaining CNN's computational efficiency. Extensive experiments on three public HSI datasets demonstrate that CGDC-Net outperforms state-of-the-art methods. Code is available at https://github.com/huanliu233/CGDC Net.