Research And Implementation Of Continuous Spectral Feature Expression Learning Algorithm For Images

Hyperspectral images play a very important role in the fields of remote sensing,medical diagnosis,and space detection.Hyperspectral images are collected by hyperspectral sensors,namely spectral imager.The spectral imager can provide tens to hundreds of subdivided narrow-band spectral information for the target area in the spectral band,and each pixel can generate a complete and continuous spectral curve.This enables substances that were not detectable in broadband remote sensing to be detected in hyperspectral.Different substances have different spectra,and specific substances have specific spectra.The more spectral channels,that is,the higher the spectral resolution,the stronger the ability to distinguish objects.However,the acquisition of hyperspectral images is more complicated,and in existing research methods,a single training model can only reconstruct hyperspectral images with a specific number of spectral channels,which limits the acquisition and application of hyperspectral images.A research hotspot in the field of computer vision is to reconstruct hyperspectral images based on convolutional neural networks.However,traditional convolutional neural network models can only reconstruct hyperspectral images with a specific number of channels,and the model reconstruction effect and generalization performance are poor.The research content of this paper is as follows:First,by studying the existing hyperspectral image reconstruction methods and the mapping relationship between hyperspectral images and RGB images,the spatial superresolution is extended to spectral super-resolution,A learning algorithm for image continuous spectral feature expression is proposed,which extracts the spatial spectral feature of the image and takes the feature,the spatial coordinates and spectral coordinates of the image as the input of the image continuous expression,and then reconstruct the hyperspectral image with variable spectral resolution,and realize the reconstruction of hyperspectral image with variable spectral channel number from RGB image;Second,applying knowledge transfer to hyperspectral image reconstruction tasks,a hyperspectral image reconstruction model based on knowledge transfer is proposed,and the reconstruction performance of the student network model is enhanced by pre-training the teacher network.The experimental results show that the variable spectral resolution image reconstruction model proposed in this paper has different improvement for performance compared with other reconstruction models.The variable spectral resolution image reconstruction model uses the encoder-decoder architecture to continuously express image features,and can reconstruct hyperspectral images with different numbers of spectral channels,expanding the spectral resolution of the reconstructed image;The hyperspectral image reconstruction model based on knowledge transfer utilizes the teacher network with stronger learning ability and generalization ability,and further improves the performance of the student network in reconstructing hyperspectral images.