Deep Learning Based Hyperspectral Image Target Detection

In recent years,deep learning has greatly improved the cognitive ability of the network by extracting deep features,and has been successfully applied in the field of feature extraction and classification of hyperspectral images.However,in the field of hyperspectral image target detection,the priori target spectral information is very scarce,which leads to the network can not be trained,and seriously hinders the construction of depth model.Aiming at this problem,this paper analyses the spectral characteristics of target from two perspectives of transfer learning and sample generation,designs two kinds of target detection frameworks for deep learning,and solves the problem of sparse training samples.The main work of this paper is as follows:Firstly,from the perspective of migration learning,combining the idea of pixel pairs,a labeled data set is used to generate a pair of pixels.According to the existing label information,new labels are allocated to each pair of pixels.A set of data sets of pixels is constructed and a network is trained to learn the spectral similarity between pixels.In the test,the sample with the greatest difference from the target spectrum is found as the background sample through the linear prediction algorithm.Then the test pixels are paired with the target and background samples respectively and sent into the network.Finally,the two groups of results are adaptively fused to get the final similarity score.This method effectively solves the problem of sample scarcity,and adds the similarity feature of depth pixels to enhance the detection ability of the network.However,this method needs to train the same sensors as the test data,so it limits the scope of application.Secondly,from the perspective of sample generation,improving the method based on transferring learning,the target sample generator is constructed by using the characteristics of output approximation input of Auto Encoder,and the low-level features are fused with the high-level features to preserve the original texture information of the spectrum.Linear prediction algorithm is used to find background samples and enhance the data.The training sample set is obtained by matching the target sample and background sample.Pixel similarity recognition network is built,and two different modes are used for experiment and analysis,and the optimal mode is selected.This method can make full use of the known information,greatly expand the samples without reference data,and solve the problem of limited range caused by different sensors in migration learning.