Hyperspectral Target Detection Based on Prior Spectral Perception and Local Graph Fusion

With the development of hyperspectral sensing technology, hyperspectral target detection technology plays an important role in remote target detection. However, existing hyperspectral target detection models are poorly adapted to complex backgrounds and mainly focus on the spectral domain, making less use of spatial structure information leading to low target detection rates. Therefore, a new target detection algorithm based on the prior spectral perception and local graph fusion (SPLGF) is proposed. Firstly, the prior spectrum-guided target extraction method is established. This method can take full advantage of the background and target spectral information by local inner and outer window linkage, reduce the impact of spectral variability on target acquisition performance, and improve detection stability. Secondly, the target enhancement strategy based on the Gabor multi-feature graph is proposed. This technique makes full use of multi-directional and multi-scale spatial information, which can reduce the influence of brightness, contrast and amplitude variation on detection performance due to light and angle. Finally, spatial-spectral fusion is executed to achieve target detection. It can make full use of spectral and spatial structure information to improve the target detection effect. Publicly available datasets and real collected datasets are adopted to check the validity of the proposed method. After comparison, it is found that the proposed algorithm has better detection effect than existing baseline methods. The maximum improvement in AUC values are 16.56%-88.16% across the eight datasets.