Research And Implementation Of Multispectral Image Reconstruction Technology Based On Deep Learning

The multispectral image reconstruction technology of a single image refers to that a 3-channel RGB image can be reconstructed into the corresponding multi-channel spectral image by using algorithms.Compared with RGB images,multispectral images contain more information and are easier to find the hidden information behind the image.Therefore,multispectral image reconstruction technology has been widely used in military target detection,video surveillance,artwork restoration and other fields.The spectral image reconstruction algorithm based on deep learning refers to learning the mapping relationship from RGB images to spectral images by establishing an end-to-end network model.Convolutional neural networks have become the mainstream framework in the field of spectral image reconstruction due to its powerful feature extraction capabilities.However,the disadvantage is that when extracting image feature information,only deepening the network layer is considered for deep feature information extraction,ignoring the importance of shallow feature information for image reconstruction and the inability of convolutional layers to maintain global feature diversity;At the same time,most reconstruction algorithms treat all pixel-level features on the image equally and ignore the differences between different feature information,which leads to the problem of low reconstruction accuracy of spectral images.In order to solve the above problems,the main research work of this paper is as follows:Firstly,a residual dense network based on attention mechanism is proposed.Aiming at the problem of different expression capabilities of feature information on different channels,a double residual channel attention module was introduced on the basis of residual dense block to improve the learning ability of network model to extract feature information between different channels.Secondly,a multi-scale residual network based on mixed attention is proposed.From the perspective of maintaining the diversity of original feature information,the network structure designs a multi-scale residual block to extract multi-scale feature information by using convolution layers with different sizes of convolution kernels to increases the local feature information extraction capability of the network;At the same time,a mixed attention module is designed,which combines the channel dimension and the pixel dimension to adaptively scale the extracted feature information,so as to improve the expression ability of important information and improve the accuracy of image reconstruction.The experimental results on the ICVL data set and the CAVE data set show that,compared with other classical and advanced algorithms,the spectral image reconstructed by the improved algorithm in this paper has richer feature information,higher reconstruction accuracy and less error.Finally,a spectral image reconstruction system is designed and implemented based on the proposed spectral reconstruction algorithm.The spectral image reconstruction system mainly includes four parts: personal information management,data management,model management and image reconstruction.The visual interface design of the software system can help users to complete the spectral image reconstruction task more intuitively and simply,and generate corresponding spectral images.