Research On Near Real Time Reconstruction Technology Of Coded Image Based On Deep Learning

Computational imaging technology[1]obtains two-dimensional spatial information of the target and one-dimensional spectral information of each pixel which finally forms a data cube of spectrum.With the application of Compressive Sensing(CS)theory in various fields,spectral imaging technology has also begun to use CS reconstruction methods,such as reedy iterative method[2]or convex optimization method[3].However,traditional CS reconstruction methods are usually compromised between spatial resolution and iteration time which make it difficult to achieve high quality and fast multi-spectral image reconstruction.Therefore,how to achieve high quality multi-spectral image reconstruction quickly has become an urgent problem in spectral imaging technology.This paper focuses on near real-time multi-spectral image reconstruction based on deep learning.The main tasks are as follows.In order to solve the problem that the sparseness of the traditional measurement matrix is not suitable for the multi-spectral image scene which makes the measurement matrix unable to use the detailed information of the multi-spectral image to achieve high-quality image reconstruction,an encoding template based on codec neural network is studied and implemented Design method.First,design the BinaryNet structure as the coding network,and use the binary weight method to train the coding network.Second,use the Multi-Layer Perceptron(MLP)network model as the decoding network to learn the nonlinear mapping function,and accurate weight Construct multi-spectral images.Finally,the multi-spectral image data is used to jointly train the codec network to train the encoding template.Through the comparison of multiple sets of simulation experiments with other measurement matrices,the same multi-spectral image reconstruction is used analyze the image evaluation indexes.The coding template proposed in this paper has higher reconstruction performance than the traditional random measurement matrix.Aiming at the problem that the traditional multi-spectral image reconstruction algorithm based on compressed sensing uses iterative methods to improve the quality of the reconstructed image which results in a long running time of the algorithm,the multi-spectral image reconstruction method based on deep learning and the residual-based multi-spectral image reconstruction method of convolutional autoencoder is studied.First,the multi-spectral data is encoded by the encoder in the convolutional auto-encoder.Then,the multi-spectral image is reconstructed by the decoder's decoding ability,and the depth residual structure is used in the decoder to further improve the decoding capability.By comparing the results of multiple simulation experiments with other methods and analyzing the image evaluation indicators,it can be seen that the methods proposed in this paper are superior to the traditional compressed sensing methods in terms of speed and the image quality.Because of the reconstruction algorithm in the snapshot compressed spectral imaging model does not use the spectral correlation between the multi-spectral image data,so the spectral correlation of the reconstructed data is not strong,and the spatial resolution is not high.A fast multi-spectral image reconstruction method based on the combination of U-net and SA-cGAN is proposed.First,based on the conditional generation adversarial network,the image restoration capability is used to achieve image reconstruction.Then,the self-attention mechanism(SA)is introduced into the GAN framework,and all spectral channels share the same attention map to maintain spectral similarity in spectral image reconstruction.Finally,U-net replaces the traditional fully connected layer to simplify the network structure.By comparing the results of multiple simulation experiments with other methods and analyzing their image evaluation indicators,the experimental results show that the fast multi-spectral image reconstruction method based on the combination of U-net and SA-cGAN proposed in this paper has a higher image correlation between quality and spectrum,and a faster reconstruction speed.