| Unmanned platforms have the advantages of low cost and strong sustainability.They can perform detection,search and rescue,security,patrol and other tasks in many fields,such as sea,land and air.Infrared multi-spectral imaging has the advantage of simultaneously detecting the spectral and geometric characteristics of targets compared with single-band imaging.Therefore,the infrared multi-spectral imaging of unmanned platform is the most potential photoelectric load at present.It has been widely used in remote sensing geological survey,forest monitoring,agricultural application and marine research.With the deepening of application,people put forward higher requirements for spatial-spectral-temporal resolution of infrared multispectral imaging.In this paper,the key technologies of infrared medium wave multispectral imaging for unmanned platform based on dual prism computational coding are studied.The main work is as follows.Aiming at the problem that traditional multi-spectral imaging technology can not take into account the high spatial and spectral resolution,an infrared medium-wave multi-spectral imaging technology for unmanned platform based on dual prism computational coding is studied and implemented.Firstly,the cyclic S matrix is designed as the encoding template to encode and modulate the multi-spectral image after illumination.Then,a prism placed in reverse is added to eliminate the dispersion displacement and an infrared array detector is used to collect the aliasing image.Finally,the multi-scale regularized Richardson-Lucy reconstruction method is used to reconstruct the aliasing image to obtain the infrared multi-spectral image.The feasibility of the optical design scheme and the high spectral-spatial resolution of the technology can be verified by experiments,which spectral resolution is 100 nm and the PSNR values are all greater than 25 d B.Aiming at the problem that the slow reconstruction speed of traditional compressed sensing reconstruction method leads to the low frame rate of imaging,the single-band and multi-spectral reconstruction method based on depth learning is deeply studied.In single-band reconstruction,the reconstruction method based on convolutional neural network(Recon Net)is improved,and a high-precision real-time image reconstruction method based on residual network is designed and implemented.Firstly,the initial reconstructed image with relatively improved quality is obtained by convex optimization constraints on the non-linear mapping.Then,the precise single-band reconstructed image is obtained by gradually narrowing the gap between the real image and the original reconstructed image through the residual network.In multi-spectral reconstruction,a fast multi-spectral image reconstruction based on self-encoder is realized.The experimental results show that the proposed method is superior to other methods in the quality and speed,and the self-coding network also has advantages in the accuracy of reconstructed image.Aiming at the problem of low energy of reconstructed single spectral band and low contrast of infrared image with high background,the super-resolution reconstruction(EDSR)based on enhanced residual network and the image detail enhancement based on deep convolution generative adversarial network(DCGAN)are deeply analyzed,and a new method of super-resolution of infrared image is designed and implemented.Firstly,the redundant convolution layer of EDSR is eliminated to reduce model parameters;secondly,the activation channel is broadened to obtain low-level features of the image;then,the weights normalization is added to reduce the training difficulty of depth network;finally,the original image is enhanced by DCGAN as a high-resolution image training improved EDSR to achieve super-resolution detail reconstruction of infrared image end-to-end.A number of experimental results show that the proposed method is more clear in outline,richer in details,less distortion and better in effect than other methods in super-resolution reconstruction of infrared images. |
PDF Full Text Request