A Research On Hyperspectral Image Super Resolutiion Based On Deep Convolutional Network

Hyperspectral remote sensing usually collects information of objects in hundreds of continuous bands on a specific electromagnetic spectrum.It collects images with very high spectral resolution,and realizes fine differentiation of different objects through its spectral characteristics.It has important application value in military and civilian fields such as mineral exploration,precision agriculture,deep space exploration,camouflage target discrimination and so on.However,Limited by the imaging mechanism and optical devices,it is difficult for existing spectral imaging equipment to directly obtain a hyperspectral image with both high spatial resolution and high spectral resolution,which will seriously affect the application of hyperspectral images.Therefore,how to improve the spatial resolution of hyperspectral image while maintaining the spectral information is a challenging and urgent task.The research background of this paper is the lunar exploration project and the Mars exploration project.In order to meet the demand for high-quality images in deep space exploration,this paper studies the hyperspectral images super resolution(HSI SR).In recent years,deep convolutional networks have shown the great advantages in computer vision.CNN-based color image super-resolution can be directly applied to HSI SR in a band-by-band or 3-band group manner,but this method lacks consideration of the correlation between the bands,resulting in spectral distortion.HSI is a cube of data.According to the characteristics of hyperspectral data,the 3D convolution-based method takes into account both spatial and spectral information,and has achieved good results,but ignores the high degree of redundancy between the spectra.The large number of 3D convolution parameters brings some difficulties to the hyper-resolution task.Based on the 3D deep convolutional network,this paper studies the super-resolution of the hyperspectral image,which can better restore the texture details of the image and reduce the amount of calculation.The main research contents and innovations of this article are as follows:(1)In order to improve the quality of super-resolution and simplify the calculation,and to solve the problems of 3D convolution ignoring the band information redundancy and many parameters,we propose a 1D-2D spatial-spectral hyperspectral single-image super-resolution algorithm.Hyperspectral images are rich in spectral information but highly redundant between bands.The 3D volume integral solution can effectively alleviate the redundancy between the bands.The problems to be solved in HSI SR focus on spatial resolution enhancement and spectral fidelity.Therefore,the spectral features and spatial features can be extracted respectively by resolving the 3D volume integral into a 1D spectral branch and a 2D spatial branch,and the spectral features are fused to the spatial features through hierarchical lateral connections.At the same time,in order to extract spatial information more fully,the spatial branch uses partial bands.By increasing the number of channels,the diversity of spatial features is increased.(2)In order to combine the local and global features of the image to generate richer texture detail information and more realistic visual effects,we propose a 1D-2D spatial-spectral hyperspectral single-image super-resolution algorithm based on attention mechanism.In the past,the way to increase the receptive field was the stacking of convolutional layers.The attention mechanism can obtain global context information by assigning different weights to different positions of the entire image,which effectively improves the quality of SR.Experiments on hyperspectral datasets demonstrate the effectiveness of the attention mechanism.What’s more,the images generated by our algorithm are more realistic from the subjective and objective evaluations.(3)In order to make full use of multi-source data,our algorithm is not only suitable for hyperspectral single-image super-resolution without reference images,but also for hyper-spectral fusion algorithms with auxiliary images such as panchromatic images.We can realize it with multi-scale feature extraction module.Limited by complex imaging methods and optical conditions,it is difficult to achieve the highest spatial resolution and spectral resolution of remote sensing data at the same time.By making full use of multi-platform and multi-sensor observation,the capability of remote sensing data acquisition can be further improved and a multi-source complementary information fusion system can be established.The experiments over widely used benchmarks on hyperspectral fusion demonstrate that the proposed method could outperform other state-of-the-art methods,both in visual quality and quantity measurements.