We deal with the problem of blind parameter estimation of signal-dependent noise from mono-component image data. Multispectral or color images can be processed in a component-wise manner. The main results obtained rest on the assumption that the image texture and noise parameters estimation problems are interdependent. A two-dimensional fractal Brownian motion (fBm) model is used for locally describing image texture. A polynomial model is assumed for the purpose of describing the signal-dependent noise variance dependence on image intensity. Using the maximum likelihood approach, estimates of both fBm-model and noise parameters are obtained. It is demonstrated that Fisher information (FI) on noise parameters contained in an image is distributed nonuniformly over intensity coordinates (an image intensity range). It is also shown how to find the most informative intensities and the corresponding image areas for a given noisy image. The proposed estimator benefits from these detected areas to improve the estimation accuracy of signal-dependent noise parameters. Finally, the potential estimation accuracy (Cramér-Rao Lower Bound, or CRLB) of noise parameters is derived, providing confidence intervals of these estimates for a given image. In the experiment, the proposed and existing state-of-the-art noise variance estimators are compared for a large image database using CRLB-based statistical efficiency criteria.

A new algorithm is described for estimating the noise model parameters in hyperspectral data when neither noise components variance nor noise spatial/spectral correlation priors are available. A maximum likelihood (ML) technique is introduced for checking the noise spatial correlation hypothesis and estimating the spatial correlation function width alongside with estimating signal-independent and signal-dependent noise components variance. The hyperspectral image is assumed to match a limited set of assumptions. A three-dimensional (3-D) fractional Brownian motion (fBm) model is introduced for describing locally the texture of the 3-D image noisy textural fragment. Nonstationarity of the useful image signal is taken into account by performing the estimation locally on a 3-D block-by-block basis. The accuracy of the proposed algorithm is first illustrated for synthetic images obtained from either pure fBm or almost noise-free AVIRIS hyperspectral images artificially degraded with spatially correlated noise. The results obtained for synthetic images demonstrate appropriate accuracy and robustness of the proposed method. Then results obtained for real life AVIRIS hyperspectral data sets confirm the noise spatial uncorrelation hypothesis for images acquired by the AVIRIS sensor. Conclusions and open problems are outlined.

The problem of automatic detection of image areas appropriate for accurate estimation of additive noise standard deviation (STD) irrespectively to processed image properties is considered in this paper. For accurate estimation of either image texture or noise STD, we distinguish two complementary informative maps: noise- (NI-) and texture- (TI-) informative ones. The NI map is determined and iteratively upgraded based on the Fisher information on noise STD calculated in scanning window (SW) fashion. Fractional Brownian motion (fBm) model for image texture is used to derive the required Fisher information. To extract final noise STD from NI map, fBm- and DCT-based estimators are implemented. The performance of these two estimators is comparatively assessed on large image database for different noise levels. It is also compared with performance of two competitive state-of-the-art estimators recently published. Utilizing NI map along with DCT-based noise STD estimator has proved to be significantly more efficient.

Detecting similarities between image patches and measuring their mutual displacement are important parts in the registration of multimodal remote sensing (RS) images. Deep learning approaches advance the discriminative power of learned similarity measures (SM). However, their ability to find the best spatial alignment of the compared patches is often ignored. We propose to unify the patch discrimination and localization problems by assuming that the more accurately two patches can be aligned, the more similar they are. The uncertainty or confidence in the localization of a patch pair serves as a similarity measure of these patches. We train a two-channel patch matching convolutional neural network (CNN), called DLSM, to solve a regression problem with uncertainty. This CNN inputs two multimodal patches, and outputs a prediction of the translation vector between the input patches as well as the uncertainty of this prediction in the form of an error covariance matrix of the translation vector. The proposed patch matching CNN predicts a normal two-dimensional distribution of the translation vector rather than a simple value of it. The determinant of the covariance matrix is used as a measure of uncertainty in the matching of patches and also as a measure of similarity between patches. For training, we used the Siamese architecture with three towers. During training, the input of two towers is the same pair of multimodal patches but shifted by a random translation; the last tower is fed by a pair of dissimilar patches. Experiments performed on a large base of real RS images show that the proposed DLSM has both a higher discriminative power and a more precise localization compared to existing hand-crafted SMs and SMs trained with conventional losses. Unlike existing SMs, DLSM correctly predicts translation error distribution ellipse for different modalities, noise level, isotropic, and anisotropic structures.

Finding putative correspondences between a pair of images is an important prerequisite for image registration. In complex cases such as multimodal registration, a true match could be less plausible than a false match within a search zone. Under these conditions, it is important to detect all plausible matches. This could be achieved by an exhaustive search using a handcrafted similarity measure (SM, e.g., mutual information). It is promising to replace handcrafted SMs with deep learning ones that offer better performance. However, the latter are not designed for an exhaustive search of all matches but for finding the most plausible one. In this paper, we propose a deep-learning-based solution for exhaustive multiple match search between two images within a predefined search area. We design a computationally efficient convolutional neural network (CNN) that takes as input a template fragment from one image, a search fragment from another image and produces an SM map covering the entire search area in spatial dimensions. This SM map finds multiple plausible matches, locates each match with subpixel accuracy and provides a covariance matrix of localization errors for each match. The proposed CNN is trained with a specially designed loss function that enforces the translation and rotation invariance of the SM map and enables the detection of matches that have no associated ground truth data (e.g., multiple matches for repetitive textures). We validate the approach on multimodal remote sensing images and show that the proposed “area” SM performs better than “point” SM.