The Guided Network Algorithm For Hyperspectral Image Super-Resolution

Hyperspectral images(HSIs)are capable of continuously capturing images of different spectral bands in the same scene,with extremely high spectral resolution and up to several tens or even hundreds of spectral bands.Due to the fact that different bands in HSIs have varying sensitivities to different substances,HSIs contain a large amount of spectral information that can be used to reflect the material characteristics of the current scene.With this highly sensitive perception,HSIs are playing an increasingly important role in many fields,such as environmental,geological,defense,and agriculture.However,existing spectral imaging devices are limited by the amount of incident energy,and obtaining high-resolution hyperspectral images(HR-HSIs)directly comes at a great cost.As a result,imaging systems usually need to balance spatial and spectral resolution,typically at the expense of spatial resolution,which severely limits the application of HSIs in real-world scenarios.Hyperspectral super-resolution(HISR)technology can effectively enhance the resolution of HSIs by fusing low-resolution HSIs(LR-HSIs)and high-resolution multispectral images(HR-MSIs)of the same scene to obtain high-resolution hyperspectral images.Therefore,it has become an effective way to improve the resolution of HSIs.HR-MSIs refer to images that include multiple spectral bands,usually only a few to tens,such as RGB images.Because HR-MSIs do not require a large amount of incident energy to disperse different spectral bands,they can have higher spatial resolution than HSIs and can be used to supplement the spatial information in HSIs.In recent years,deep neural networks,particularly Convolutional NeuralNetworks(CNNs),have emerged as a core technology for high-resolution hyperspectral image superresolution.They have shown promising results,but existing CNN-based methods for hyperspectral image super-resolution still have some limitations.Firstly,these methods often have complex network structures and a large number of parameters,which significantly consume computational resources and require long training and execution times.Secondly,existing methods generally use the entire multispectral image for single-scale fusion,which makes it difficult for the network to accurately reconstruct images in cases where there is a significant difference in clarity between HR-MSI and LR-HSI.Thirdly,these methods have limited generalization ability,and it is difficult to extend them to other image super-resolution problems,such as panchromatic sharpening or single-image super-resolution.To overcome this limitations,this paper proposes a high-resolution-guided general CNN fusion framework,GuidedNet.The framework consists of two branches: 1)a highresolution guidance branch that decomposes the high-resolution guidance image into multiple scales; and 2)a feature reconstruction branch that fuses the low-resolution image with the multi-scale high-resolution guidance images obtained from the high-resolution guidance branch to reconstruct a high-resolution fused image.GuidedNet can use residual learning to simultaneously improve spatial quality and maintain spectral information.Additionally,the proposed framework is implemented using a recursive and progressive strategy,which can improve performance while significantly reducing the number of network parameters,and ensures network stability by supervising multiple intermediate outputs.Thanks to these charac-teristics,the proposed method achieves good super-resolution performance,and the main contributions of this paper can be summarized as follows:1)A general CNN fusion framework,GuidedNet,was proposed and successfully applied to multiple image resolution enhancement problems,such as hyperspectral image super-resolution,pan-sharpening,and single-image super-resolution,achieving satisfactory results in each task.2)Two novel network branches were designed,namely the high-resolution guidance branch and the feature reconstruction branch,to exploit the multi-scale information of the high-resolution guidance image(HR-MSI)and reconstruct the fused high-resolution output.They feature multi-scale information fusion,progressive feature injection,and progressive feature reconstruction.The multi-scale framework can more accurately capture rich structural information and better adapt to the large super-resolution scale reconstruction task compared with single-scale methods.In addition,GuidedNet supervises the intermediate results to improve stability and obtain satisfactory results.3)Compared with previously developed methods,GuidedNet achieves superior performance and has fewer network parameters due to the use of recursive structures.Most importantly,it exhibits strong generalization capabilities for reconstruction at different super-resolution scales and good adaptability to other image resolution enhancement tasks.