Hyperspectral Image Target Detection Technology Integrating Domain Knowledge And Deep Network

Deep network has been well applied in the field of hyperspectral image classification and segmentation with its excellent feature extraction capabilities and excellent data analysis capabilities.However,target samples are extremely scarce,lack domain knowledge guidance and complex and diverse scenarios lead to problems such as poor separability of positive samples and negative samples,which severely limit the application of deep network in the field of target detection.The domain knowledge of hyperspectral images is mainly embodied in prior knowledge,spatial neighborhood information and logical relations.Making full use of the domain knowledge of hyperspectral images can reduce the dependence of deep network on data labels and help improve the detection accuracy of deep network models.Therefore,this paper focuses on solving the problem of small samples,using data enhancement from data space and feature space,and fully mining the domain knowledge of hyperspectral images to assist in improving the target detection performance of hyperspectral images under the problem of small samples.The main research contents are as follows:（1）Generate data from the data space.Build a data generation model（UAE-Net）for the problem of insufficient target sample size.Concatenate the hidden layer features after hyperspectral data encoding and the high-level features of the decoding stage through feature cascade,effectively capturing the contextual information of hyperspectral images.The addition of the sawtooth dilated convolution aggregates multiple proportions of spatial information for the residual network,realizes that multi-level and multi-scale information complements each other,and enhances the refined representation of details such as the edge contour of the ground feature.It also enables the network to have more complete semantic expression capabilities.Experimental results show that the residual network target detection model based on prior data generation can effectively improve the target detection performance of hyperspectral images for small sample problems.（2）Data enhancement is carried out from the perspective of feature space.Aiming at the problem of insufficient utilization of domain knowledge in hyperspectral image target detection,this paper uses paired combination of positive and negative samples to achieve the expansion of data in the feature space.Tag attributes are an important data-assisted feature.Through hierarchical,diversified,and differentiated representations,logical relationships between tag features of different categories are established.The data after sample expansion and balanced processing of positive and negative feature information provides the deep network with stable and rich representative features that contain domain knowledge.Use HTD-Net to learn the subtle differences between spectra and fuse decision information with the aid of prior knowledge.Experimental results show that the prior knowledge-assisted target detection model based on feature enhancement proposed in this paper can effectively improve the performance of hyperspectral image target detection under weak prior knowledge.