Research On Hyperspectral Images Visualization Method Based On Deep Learning

The ground data processing system needs the visualization of hyperspectral images for fast viewing and information measurement.Compared with multispectral images,hyperspectral images have more abundant information,which provides favorable conditions for hyperspectral image visualization.Traditional hyperspectral image visualization algorithms usually have the problems of color distortion and detail loss.This paper analyzes the problems existing in the existing algorithms,combines the advantages of neural network technology in feature extraction and image reconstruction,and uses neural network technology to model the problem of hyperspectral image visualization.The main research contents are as follows1.For traditional hyperspectral image visualization,the method of band selection or dimension reduction of hyperspectral images can not produce stable details and good color.To solve this problem,this paper proposes a hyperspectral image fusion method based on multi band selection and supervised learning(CNNHV).In the training phase,multispectral image is introduced as the reference image,and convolution neural network is used to model the fusion of hyperspectral image and multispectral image.The experimental results show that,compared with the five latest hyperspectral visualization methods,this method achieves the highest evaluation on the spectral angle,relative average spectral error and other color fidelity metrics.2.The proposed CNNHV method is usually only suitable for hyperspectral images with visible wavelength range.For the visualization of near-infrared hyperspectral images,it is difficult to produce the same reasonable color as the corresponding multispectral images using this method.This problem also exists in the traditional hyperspectral image visualization algorithm.Therefore,this paper proposes an end-to-end hyperspectral image visualization method based on Generative Adversarial Networks(GANHV).This method can fuse near-infrared hyperspectral images into three bands.At the same time,the discriminant network is introduced to construct the generation confrontation network to improve the performance of the proposed method.Experimental results show that,compared with the five latest hyperspectral visualization methods,the proposed method achieves the highest evaluation on the color fidelity metrics.3.Compared with the direct acquisition of hyperspectral images,the hyperspectral image generation method based on multispectral image dimension elevation can save a lot of acquisition,transmission and storage costs.At the same time,the combination of multispectral image dimension elevation method and hyperspectral image visualization method can realize the fast transmission and lightweight storage of hyperspectral image.In the process of dimension elevation from multispectral image to hyperspectral image,a large number of bands need to be reconstructed.However,the lack of data characteristics makes this process very difficult.In order to verify the feasibility of multispectral image dimensionality enhancement,this paper proposes to use convolution neural network to transform three band color image into hyperspectral image with hundreds of bands.In order to intuitively evaluate the quality of dimensionality enhancement results,four visualization algorithms are used to test their visualization effects.The visual results show that the method can generate images similar to real hyperspectral images.In addition,in order to test the reconstruction effect of the proposed method in different wavelength ranges,six objective indexes are used to evaluate the structure and spectral loss of the dimension enhancement results.The evaluation results show that,compared with the near-infrared wavelength range,this method achieves higher structural similarity and lower spectral information loss in the visible wavelength range.To sum up,the two visualization methods proposed in this paper have good performance in maintaining the structure and color.At the same time,the proposed method can generate images similar to the real hyperspectral images.