Semi-supervised Hyperspectral Image Classification Based On Generative Adversarial Network

A kind of new detection mode is proposed for the preparation results of various fluorescent molecule samples in combination with the microscopic hyperspectral imaging system and generative adversarial network(GAN).Compared with the traditional ultraviolet-excited imaging,the algorithm proposed in this paper remedies the shortage of controls details difficultly and need more expert knowledge,and enhances the efficiency and accuracy of judging the preparation results of various fluorescent moleculars.Microscopic hyperspectral system is constructed at first in this paper for the data acquisition of various fluorescent molecular and for the data preprocessing.The optimal imaging model is explored and analyzed,and the relevant parameters and spectral characteristics of dataset are also analyzed.In the pretreatment process,the absorbance conversion is adopted for spectral correction,adaptive band selection algorithm(ABS)is adopted for band selection,and standardized processing is adopted to eliminate the influence of data dimension to improve the network convergence rate.Later,a kind of convolution-based semi-supervised generative adversarial network model is proposed.This algorithm inputs the authentic specimens without label into model for feature extraction,and fully utilizes the authentic specimens of data concentration to solve the difficult of less label samples in the classification of hyperspectral image.Experimental verification results show that the overall classification,average classification precision of various categories and Kappa coefficient of algorithms in this paper are obviously higher than those of traditional machine learning classification algorithm and existing semi-supervised generative adversarial network algorithm,but have poor performance in individual sample category because the absorption peak of substance exceeds the acquisition scope of hyperspectral imager.Therefore,it is difficult to distinguish the substance and background.Subsequently,algorithm would be improved against such problem.In the network training process,a kind of three-stage training method is designed in this paper.At the first stage,only authentic specimens are input into discriminator for training;At the second stage,the mean value and variance of Batch Normalization(BN)layer are fixed,and the false samples are input into discriminator for training together;At the third stage,the parameters of generator are fixed to continue training of discriminator.Three-stage training method for the design in this paper enhances the classification accuracy of network to a certain extent,and accelerate the network convergence speed at the same time.Finally,a kind of convolution-based semi-supervised generative adversarial network classification model based on space-spectrum joint feature is proposed.The input data of such model contains the voxel block of spatial information.The texture information in spatial neighborhood is extracted and different kinds of fluorescent molecules are assisted for classification to improve the classification accuracy.Experimental results show that such model effectively solves the poor classification results of individual categories in convolution-based semi-supervised generative adversarial network.