Hyperspectral Images Noise Variance Estimation Based On Unsupervised Learning

Hyperspectral images(HSIs)have noise due to photoelectric transmission.These noises not only affect people's visual quality,but also affect the quality of images.At the same time,noise assessment can not only be used as an evaluation index of sensor quality,but also an important parameter for post-processing of HSIs.Therefore,noise variance estimation of HSIs is of great significance.Based on unsupervised learning theory,this paper estimates the noise variance of image by constructing a non-linear model,and improves the accuracy of noise variance estimation by utilizing spatial correlation of objects and spectral correlation of HSIs.The main contents of this paper are as follows:(1)In this paper,a spatial spectral principal component analysis(SSPCA)algorithm is proposed.Firstly,the reconstructed image is reconstructed by rearranging the spatial spectral dimension of the original image.Then,the noise and signal are separated by using the principal component analysis algorithm.Finally,the noise variance of the image is obtained.In this paper,SSPCA and two classical algorithms are compared and analyzed on simulated images and real images respectively.Although SSPCA has higher accuracy estimation results on the whole,it still does not accurately estimate noise variance in some bands.(2)In this paper,unsupervised deep learning algorithm is used for noise assessment of HSIs.Different from the common deep learning algorithm,which requires a lot of data to train the model,this method only needs one image to extract information in training,and finally obtains the estimation of noise variance of the image,which greatly reduces the training time.(3)In order to train the U-Net model more accurately and quickly,this paper optimizes and improves the structure of U-Net model.First,the activation function of the last layer of the original U-Net algorithm is changed to sigmoid.Secondly,in order to improve the training speed,the down-sampling and up-sampling layers of the original U-Net model are improved,and the number of layers of the model is reduced from 19 to 10.Then the improved U-Net model is obtained.Finally,in order to improve the accuracy of extracting information,a non-local block structure is added to the improved U-Net model,which is conducive to more accurate estimation of noise variance in images.Through experimental analysis,nonlocal U-Net has the advantage of faster and more accurate estimation of noise variance.For the two noise variance estimation algorithms proposed in this paper,experimental analysis is carried out through real images.The experimental results show that the algorithm based on unsupervised deep learning has higher accuracy and stability.