Research On Remote Sensing Image Fusion Algorithm Based On Residual Network And Attention Mechanism

Since the launch of artificial earth satellites,remote sensing technology has played an important role in economic,political,military and other fields.With the development of remote sensing technology,existing remote sensing satellites can obtain images with higher and higher spatial,temporal and spectral resolutions.However,due to technical conditions and hardware limitations,remote sensing satellites usually can only provide two types of remote sensing images : one is a panchromatic(PAN)image with high spatial resolution;the other is a multispectral(MS)image.PAN images have only one spectral channel and cannot express RGB colors,on the contrary,MS images have high color expressive ability.Therefore,remote sensing image fusion of PAN and MS is proposed to generate MS images with high resolution.At present,deep learning methods have been applied to remote sensing image fusion,and various deep learning models have achieved better performance in the field of remote sensing image fusion due to their powerful high-quality image generation capabilities.This paper focuses on designing a fusion network based on deep learning to improve the spatial resolution of the fused image while maintaining its spectral quality.This paper mainly carried out the following work:1.A remote sensing image fusion method based on multi-scale convolution and residual channel attention is proposed.The model in this paper designs dual-stream input PAN and MS images,uses multi-scale convolutional layers to adaptively extract spectral information and spatial details,and residual connections enable the network to adapt to deeper network structures and avoid network degradation problems.In the fusion process,the residual channel attention mechanism module is introduced to make the network pay more attention to the key information in the feature image.This method can focus on extracting the deep features of the image and comprehensively reconstruct the fused image.The experimental results show that the proposed method is significantly improved compared with other methods on Gao Fen-2 and World View-2 data sets.2.A remote sensing image fusion method based on multi-branch residual fusion network is proposed.Sentinel-2 involves 13 spectral bands,including four 10 m bands,six 20 m bands and three 60 m bands,which aims to improve the low resolution bands of20 m and 60 m to the highest resolution of 10 m.This fusion problem involves four 10 m high-resolution bands,which is different from the fusion problem involving only one high-resolution PAN band and is more complex.This chapter proposes a multi-branch residual fusion network,in which multiple branch networks extract features from high and low resolution band images.The subsequent network fuses them together to form a compact channel representation,represents spatial and spectral information at the same time,introduces the residual channel attention mechanism,adaptively recalibrates the feature graph and assigns more weights to the deep features.Finally,the high-resolution image is restored from the fusion features through the image reconstruction network.The Sentinel-2 data set is used in the experiment,and the proposed method can effectively fuse 10 m and 20m/60 m band images,improve the spatial resolution of the band,and has more advantages than the existing techniques.