LSDETR: An Improved DETR for Hyperspectral Image Object Detection with Enhanced Pixel Feature Preservation and Global Visual Field

The detection of hyperspectral images has a wide range of applications in food safety, aerospace and other fields. However, the traditional methods have problems such as inaccurate detection, and it is easy to be disturbed by noise. To this end, this paper proposes a hyperspectral image detection model based on DETR. The improved LSNet module is used in the feature extraction network to ensure that the pixel information is not lost, and the model has the global visual field perception ability at the same time. The sparse attention mechanism was introduced into the encoder to improve the computational accuracy and speed of the model. Finally, one-to-many matching was introduced to simplify the architecture of the Transformer decoder. Our model is trained and evaluated with SPOD dataset and SanDiego dataset and achieves good results. Our experimental results show that LSDETR is highly competitive in the field of object detection in hyperspectral images.