Research On Key Technologies Of LWIR Multispectral Computational Imaging

Long wave infrared(LWIR)is between the visible and microwave spectral bands,and its is the thermal radiation mainly emitted from the scene and target in space.Infrared multispectral imaging integrates the information representation of spatial and spectral dimensions.Compared with single spatial information,it has obvious advantages in image fine identification,scene temperature inversion and material structure analysis.It is widely used in forest fire prevention and environmental monitoring,and has a wide range of applications.Traditional multispectral imaging components need to design imaging components independently for multiple spectral channels,which leads to low utilization of energy and bandwidth,making it difficult to take into account the spatial spectral temporal resolution.With the development of hyperspectral imaging theory and the advantages of space-time compression coding,hyperspectral imaging has become a research hotspot.This paper focuses on the key technology of LWIR multispectral reconstruction based on computational imaging.The main work is as follows.(1)In order to solve the problems of poor real-time performance of existing deep learning compressed spectral reconstruction methods,large consumption of computing resources and insufficient mining of compressed sampling image structure information,this paper studies and implements an infrared compressed spectral image reconstruction method based on multi-scale self-attention mechanism.A novel network model for infrared compressed spectral image reconstruction is designed by analyzing the relevant theoretical models of spectral coding signal acquisition and signal reconstruction.The codec infrastructure is used to solve the convex optimization problem in image decoding,and the original codec structure is improved based on the multi-scale self-attention modular structure,with the emphasis on enhancing the hidden information hiding in the network model In order to improve the accuracy of spatial spectral reconstruction.The network model is pretrained on the visible open spectrum image dataset,and the related data are tested.At the same time,based on the self-built LWIR multispectral image dataset,transfer learning correlation method is used to fine tune the pretrained model to adapt to the reconstruction of LWIR compressed spectrum.The experimental result shows that the proposed method can effectively reconstruct the infrared coded spectral image,and accurately restore the spatial and spectral information of the multispectral image in near real time with less memory consumption.(2)Aiming at the problem of low spatial resolution of reconstructed image caused by energy attenuation of long wave infrared coding template,this paper studies and implements a super-resolution method of long wave infrared single frame image based on two-stage expansion model.Based on the traditional MAP iterative algorithm,this paper analyzes the spatial degradation representation model of the image,designs the depth expansion linear network model with reference to its calculation mode,studies and designs the parameter learning module and the improved de coder model,and makes a linear design of the l2 norm convex optimization structure in the iterative model.Based on the self-built long wave infrared high-resolution image dataset,the training of the model is completed.Through the test of several groups of single frame and multi spectral images randomly selected,the advantages of this method in detail enhancement of single frame image are verified,and the quality of reconstructed spectral image is improved to a certain extent.(3)Aiming at the problem of infrared image contrast reduction caused by the target energy loss in the process of coding and signal reconstruction,this paper studies and realizes the infrared image contrast enhancement method based on improved U-Net network.Through the process of image contrast degradation,an image contrast degradation model is established.Based on the degradation model,a method of infrared image contrast enhancement using signal restoration model is studied and realized.Using U-Net model as the basic architecture,the spatial resampling structure is optimized,and the original multiscale intermediate feature output is replaced by intermediate output of different scale target images.The original interpolation up sampling mode is improved by using the features between the joint encoder and adjacent levels,and the spatial feature up sampling with higher accuracy is realized.In order to reduce the difference between the optimization results based on the image restoration model and the subjective visual effect,the perceptual loss combined with l2 loss is used to balance the differences among the image detail performance,visual perception and contrast perception.The model is trained on the self-built infrared multispectral dataset,and tested on several groups of randomly selected single frame and single band infrared long wave image data.The experimental results show that the method studied in this paper maintains good detail information and visual optimization effect while image enhancement.