DCH-Net: A hyperspectral object detection network with differential convolution and spectral gradient fusion

Hyperspectral remote sensing images provide rich spectral information, making them widely used in downstream tasks such as object detection. However, existing detection methods often struggle to capture local variations along the spectral dimension, limiting the full exploitation of spectral information. To address this issue, we propose Differential Convolution-based Hyperspectral detection Network (DCH-Net) to effectively exploit the differential information in hyperspectral images. First, we design a Dual Path Diverse Mixture of Experts module to dynamically integrate the original channels with differential information. Second, we develop an Adaptive Deformable Attention mechanism that incorporates the image content of reference points into sampling position prediction, enabling effective use of local contextual information. In addition, considering the limited quantity and scene coverage of existing hyperspectral object detection datasets, we construct a hyperspectral ship detection dataset named SHD-25. A spectral-correlation-based target enhancement strategy is also introduced to further improve small target detectability. Experimental results demonstrate that our method outperforms existing approaches across multiple benchmarks, indicating its effectiveness for hyperspectral object detection. The source code and dataset will be publicly available at https://github.com/nal793/DCH-Net.