Processing of remote sensing data is often based on an assumption that noise parameters in component images are a priori known. If this assumption is not valid, it is desired to perform estimation of noise parameters directly from noisy image patches. In this paper, two estimators - model- and learning-based ones possessing the ability to evaluate noise standard deviation (SD) or variance and to predict estimation accuracy for each image patch are considered. The former approach is the representative of maximum likelihood estimator (MLE) of parameters for anisotropic fractional Brownian motion (afBm) field whilst the learning-based one is the representative of convolutional neural networks (CNN) that employs training on real-life images. Our goal is to compare the performance for two cases: for pure afBm data and for real-life images. It is shown that the learning-based approach occurs to be less effective for pure afBm data since it produces a certain bias whilst the model-based approach runs into problems for complex image patches in reallife images. Based on this analysis, we propose to use synthetic afBm data as additional source of training data for learning-based methods of noise parameters estimation. By mixing real and synthetic data for training of the NoiseNet CNN, we were able to improve its performance in both domains. For afBm data, NoiseNet bias was significantly reduced and ability to predict noise SD estimates confidence improved. On NED2012 database of real images, the modified NoiseNet reduces signal-independent noise SD component estimation error by about 40% as compared to the original CNN version.
This paper investigates the problem of blind noise parameters estimation (BNPE) of multispectral/hyperspectral images (M/HSI) using deep convolutional neural networks. In contrast to single-band images, M/HSI possess property of interband correlation that can effectively improve quality and accuracy of BNPE. Therefore, in this work, we extend a previously proposed CNN for BNPE of single-band images, called NoiseNet, to vector case. The proposed vNoiseNet Convolutional Neural Network (CNN) can be applied to three-band images including RGB images from Digital Single- Lens Reflex (DSLR) cameras, and subsets of M/HSI bands. Training data for the proposed CNN were obtained from three sources: calibrated images captured by DSLR Nikon D80 camera, AVIRIS data with accurately estimated noise parameters, and Sentinel-2 data with synthetic noise. The vNoiseNet estimates both sub-band (component) image noise variance and uncertainty of this estimate from 32×32×3 image patches. On the basis of a set of such estimates, both signal-independent (SI) and signal-dependent (SD) noise component parameters can be robustly estimated. Experiments on NED2012 database, AVIRIS data and Sentinel-2 data demonstrate high accuracy of the proposed CNN.
Image restoration is a necessary stage in the processing of remotely sensed hyperspectral images, when they are severely degraded by blur and noise. We address the semi-blind restoration of such degraded images component-wise, according to a sequential scheme. By semi-blind, we mean introducing a minimum of a priori knowledge on main unknowns in the restoration process. For each degraded component image, main unknowns are the point spread function of the blur, the original component image and the noise level. Then, the sequential component-wise scheme amounts in a first stage to estimating the blur point spread function directly from the considered degraded component image and in a second and final stage, deconvolving the degraded channel by using the PSF previously estimated. Our contribution is to improve further the sequential component-wise semi-blind variants of a recently proposed method. In this work, modifications previously introduced separately are applied all together. All these modifications together are beneficial as they tend to make the newly proposed method as independent as possible of the data content and their degradations. The resulting method is experimentally compared against its original version and the best ADMM-based alternative found experimentally in previous works. The tests are performed on three real Specim-AISA-Eagle hyperspectral images. The component images of these images are degraded synthetically with eight real and arbitrary blurs. Our attention is mainly paid to the objective analysis of the l1-norm of the estimation errors. Experimental results of this comparative analysis show that the newly proposed method exhibits interesting competitive performances and can outperform the methods involved in the experimental comparison.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.