A Study Of Multiband Image Adaptive Fusion Based On Deep Learning

Multi-band image fusion is one of the key technologies of intelligent detection systems.Because most current image fusion methods use two images as fusion objects and rely on prior knowledge to select image decomposition filters,decomposition layers,and fusion rules,the fusion effect is very uncertain,and it cannot adapt to current intelligent detection systems.The need for adaptive fusion,so this paper uses a deep neural network to automatically learn the appropriate model for multi-band image adaptive fusion:(1)An adaptive method for multi-band image fusion decomposition and reconstruction filter is proposed.For multi-scale transform(MST)fusion method,the problem of selecting the appropriate decomposition filter based on prior knowledge is required.On the basis of the training basic unit,the basic unit is stacked to build a deep-stacked convolutional neural network(DSCNN),adjusts the DSCNN end-to-end using the same data set.The multi-band image can be decomposed into high-frequency images and low-frequency images by DSCNN adaptively,and the decomposition and reconstruction filter adaptive in multi-band image fusion process is realized.(2)A multi-band image fusion rule adaptation method is proposed.Aiming at the problem that the traditional fusion method relies on prior knowledge to select the proper fusion rules,the adaptive algorithm can be implemented by training the Deep Gate Convolutional Neural Network(DGCNN)to establish the appropriate data set for multi-band image fusion..The multi-band image is input into the DGCNN by DSCNN decomposition results,and the two together realize the multi-band image adaptive fusion.(3)A multi-band fusion image adaptive colorization approach is proposed.For the problem that grayscale fusion results are not conducive to human eye observation,by building a suitable data set and training Deep Convolutional Neural Network(DCNN)to adaptively generate color information according to the input grayscale image,multiple Band fusion image adaptive colorization.