V. V. Lukin

Image DCT coefficient statistics and their use in blind noise variance estimation

Statistical characteristics of DCT coefficients in 8×8 blocks of noise-free and noisy images are considered. It is shown that DCT coefficients obey generalized Gaussian distribution where its parameters are individual for different DCT coefficients and depend upon image and noise properties. Analysis of statistics is carried out for all image blocks and separately for blocks recognized as homogeneous. Several approaches to robust scale estimation of DCT coefficients are discussed and the corresponding techniques for blind estimation of noise standard deviation (variance) are proposed and studied. It is demonstrated that joint processing of estimates obtained from a set of homogeneous blocks produces less biased estimation.

Performance analys of segmentation-based method for blind evaluation of additive noise in images

There is a practical need in blind and automatic evaluation of noise characteristics in remote sensing images, especially multi- and hyperspectral ones. It is both desirable to evaluate noise type and variance [1]. This is needed for image filtering, classification, compression, edge detection and other operations of image processing. In the paper we address the task of noise variance evaluation supposing that noise type is known a priori or determined in advance [1]. The main focus is on estimation of additive noise variance although the obtained results can be easily generalized to other types of noise (signal-dependent, multiplicative).

Filtering of frequency modulated signals embedded in α-stable noise using robust DFT forms

A task of frequency modulated (FM) signal filtering by means of robust DFT is considered. It is shown that in the case of non-Gaussian noise with symmetric heavy-tailed pdf a standard DFT cannot efficiently suppress outliers that contaminate the signal. Three robust DFT forms are described and the procedure adaptive to tail behavior based on the myriad estimate for deriving real and imaginary spectral components is proposed. It is demonstrated that the use of the described and proposed approaches leads to considerable improvement of signal estimate in comparison with techniques based on standard DFT. Such improvement (MSE reduction) is provided for a wide range of parameters of noise with symmetric α-stable distribution.

Meridian estimator performance for samples of generalized Gaussian distribution

Recently, a sample meridian estimator of location parameter (LP) has been proposed. It has been shown to be robust and adjustable by means of tunable parameter δ. Basic properties of this estimator were studied; however its practical behavior needs further analysis. In this paper, we study and analyze the efficiency of the meridian estimator for samples with essentially different properties. For this purpose, we use generalized Gaussian distribution as a model. Also, the efficiency of the meridian estimator is compared to the efficiency of median and myriad estimators.

Edge detection efficiency in single-look SAR images by elementary and neural network based detectors

Synthetic aperture radars (SARs) are known to be extremely useful in many applications of modern remote sensing (RS) [1]. However, an essential drawback of images acquired by SARs is the presence of intensive noise-like phenomenon called speckle which has multiplicative nature and non-Gaussian distribution law [1]. While single-look imaging mode provides the best spatial resolution, it is simultaneously characterized by the most intensive speckle with obvious non-Gaussianity. This causes problems in solving many typical image processing tasks as pre-filtering, object detection, and segmentation. Heterogeneity (in particular, edge) detection is a standard operation employed for the aforementioned tasks [1, 2]. Thus, efficient detection of edges can serve as a good pre-requisite for further processing of SAR images.

Denoising efficiency for multichannel images corrupted by signal-dependent noise

Essential improvements in quality of original images formed by multichannel (multi- and hyperspectral) sensors have been gained in recent years. In particular, level of thermal noise in acquired images has been sufficiently reduced [1]. However, there are still component (sub-band) images in obtained data for which noise level is quite high [2, 3]. One more peculiarity is that signal-dependent noise component is characterized by dominant contribution [3] for new generation of sensors. Sometimes, the component images with the lowest signal-to-noise ratio (SNR) are ignored at stages of multichannel image classification and interpreting [1, 2]. However, recent studies have demonstrated that useful information can be extracted from “noisy” sub-band images under condition that noise is reduced by an efficient pre-filtering technique [2]. Thus, an actual task is to design such efficient techniques able to cope with signal-dependent noise and to analyze their performance.

Non-parametric detection scheme for isotropic image texture with normal increments

We analyze applicability of 2D fractal Brownian motion (fBm) for real-life image textures with respect to two general fBm properties: isotropy and normality of its increments. A non-parametric detection scheme for texture satisfying these two properties is proposed. It is based on Lilliefors test for texture increments normality and Kolmogorov-Smirnov two samples test for equality of distributions of pairs of increments. The scheme is tested against large real-life images database and is shown to detect and remove such image patterns as edges, areas with clipping effects, irregular and anisotropic textures.

Compression ratio prediction for DCT-based coder

The paper deals with considering a problem typical for lossy compression of remote sensing images. While compressing an image under interest by DCT-based coders which are rather efficient it is desirable to predict what CR will be attained for a given quantization step. We show that such prediction is possible and it can be done easily, quickly, and quite accurately. Moreover, prediction can be done for practically noise-free images and images corrupted by unknown type of noise with unknown characteristics. Influence of noise type and image properties is briefly studied.

Texture region detection by trained neural network

In this paper we consider an important practical aspect of texture region detection in remote sensing images. One specific feature of our study is that we assume a processed image noisy with a priori known type and parameters of the noise. Another specific feature is that we try to detect textural regions for a wide variety of textures without having a priori knowledge of their properties. The considered task is solved by means of trained neural networks. In the paper, we analyze the aspects of choosing input local parameters used in detection (recognition) and carrying out training. The verification results provide valuable conclusions for these aspects.

The minimum number of scanning windows required for effective maximum likelihood estimation of image texture parameters and additive noise variance

In this paper, we dealt with the problem of noise variance estimation from additive mixture of the noise and an underlying image texture. Assuming fBm-model for image texture, the number Me(H,SNR)of SWs has been obtained such that statistical efficiency e of the previously designed ML noise variance estimator is close to a predefined level e = 0.9. The value Me defines a boundary between asymptotic and non-asymptotic modes of the ML estimator with respect to image fragment size (number of SWs available). For fixed SNR, Me takes minimum values for smooth textures ( H close to 0.8) and increases fast as H approaches 0. As a function of SNR, Me has minimum at approximately SNR = 1.5 and increases fast as SNR deviates from this value. These results are useful for establishing the area of applicability of noise variance estimators and to assure the quality of estimates obtained from an image texture of a given size, roughness and SNR.