Hyperspectral Ship Detection Based On Spatial-spectral Information

Hyperspectral image has the characteristics of high spectral resolution and is the ”unification of spatial and spectral image”.It contains abundant spectral information and spatial information,which can greatly facilitate the detection and recognition of small targets,hidden targets and camouflaged targets.Based on the characteristics of hyperspectral image,this thesis proposes a hyperspectral ship detection framework that ”from coarse to fine” and”combining the special and spectral information”.We have further studied each content,including sea-land segmentation,anomaly detection,band selection and ship targets detection,of the algorithm framework as follows:First,a quick adaptive threshold see-land segmentation algorithm based on sea-area constraint is proposed.The algorithm takes full advantage of the different materials in composition of sea and land as well as differences in their spectral reflectance.We increase the consistency constraint on similar seawater class,and reduce the constraint on various land materials.The proposed method can effectively reduce the segmentation noise and sharpen the segmentation edges.Second,we propose an algorithm based on low-rank tensor recovery and sparse for hyperspectral anomaly detection.As we know,spectral information is not always correct,thus some local background will be mistaken as anomalies.To solve such problem,the interaction between spatial and spectral information of hyperspectral image is fully employed.We design third-order tensor model,and then we recover low-rank background data,which leads to the sparse anomalies.In conclusion,this method can effectively reduce detection false alarm rate.Third,a hyperspectral band selection criterion based on high-order mutual information is designed for ship targets conformation.The criterion can quickly find the relationship between the complementary and redundancy of multiple bands,select the task-specific features for the purpose of discrimination.This method can explore valid part of original hyperspectral data,which means dimension reduction,while preserving its physical interpretability.Thus,it can improve the efficiency and accuracy of subsequent detection algorithm.Last,we propose an algorithm for ship target detection based on spatial-spectral feature sparse coding.The algorithm takes the spectral tensor composed of the observed spectrum and its spatial neighborhood spectra as input,then optimizes its projection direction and sparse representation coefficients.The process is equivalent to do sparse coding in the latent space of original tensor.The algorithm achieves the compressed expression of the original spectral tensor by feature projection.The sparse coding reduces the influence of ”same material with different spectra” on the test result.By using spatial neighboring information,we reduce the effect of ”different materials with same spectrum” for the test result.As a result,the proposed method could eliminate the false alarm precisely and improve the accuracy of ship detection.