Deep Feature Boosting Neural Network For Pan-sharpening

Nowadays,many fields such as agriculture,commerce,national defense and military require clear remote sensing images.However,due to the limitations of technology and transmission,a single sensor cannot contain both hyperspectral information and high spatial structure information.As a result,the remote sensing images obtained are often in Some aspects are missing,so in recent years,remote sensing image fusion technologies such as space-spectrum fusion have become research hotspots.With the development of computer and the proposal of various research theories by scholars,deep neural networks have demonstrated their dominance in the field of computer vision.More and more researchers are also applying deep neural networks in the field of Pan-sharpening,and have obtained considerable results.However,the existing Pan-sharpening methods generally have the limitation of weak feature extraction ability,and some methods do not integrate the domain knowledge of Pan-sharpening into the neural network for feature extraction.The main research of this thesis is to use the knowledge of deep neural network to build a neural network that is better at enhancing features and improve the quality of fused images suitable for Pan-sharpening.(1)Study the Enhancement Neural Network for Pan-sharpening.By the improvement of the input and the network,the features are enhanced from many aspects to improve the fusion result.The network mainly extracts sharpened panchromatic images and up-sampled multi-spectral images with dual-stream sub-networks,and then uses group module to deepen the depth of the network,improve feature quality,and combine the improved features with the shortcut.Propose spectral fidelity terms and spatial fidelity terms through the variational method to improve the spatial and spectral fidelity of the fused image.Experiments show that the Enhancement Neural Network for Pansharpening can effectively improve the quality of the fusion image.(2)Research the Extration and Excitation Neural Network for Pan-sharpening,and the feature of the network is further enhanced on the basis of the previous research.Through the research on the channel attention mechanism,combined with the dense block,a module that combines the feature enhancement and feature selection functions is proposed.While enhancing the feature,the selection function provided by the channel attention is used to complete the feature enhancement.The reduced-resolution experiment and the full-resolution experiment prove that this method can generate high-quality fusion images,and is more competitive to the other methods in terms of objective indicators and subjective observation.