Convolution-augmented transformer network for hyperspectral image subspace clustering

With the widespread and successful application of hyperspectral imaging, the task of classifying hyperspectral images has become a meaningful endeavor. This paper introduces a new approach to clustering hyperspectral images called CTNSC, which incorporates an attention mechanism to improve performance. In comparison to many current convolution-based methods for deep subspace clustering, our approach incorporates an attention mechanism to effectively capture the extensive spatial relationships that exist between objects in hyperspectral images. We combine convolution and attention through the use of Conformer blocks, enabling the network to simultaneously capture both local and global features of objects. As a result, our network can identify a more optimal deep affinity matrix, which can be utilized for spectral clustering to achieve better clustering results. We conducted experiments on three real-world datasets, and the results showed that CTNSC achieved excellent clustering performance compared to many commonly used clustering methods.