An unsupervised approach to automatically and objectively detect all classes in hyperspectral images based on physical characteristics provided by the sensors is proposed. The partitioning is done in two steps: first, the main classes are detected; then, each main class is subdivided into fine classes. The different evaluations of the proposed method on hyperspectral images show the relevance and the coherence of the partitioning results obtained in both steps.

A new blind restoration method for hyperspectral images is proposed in this paper. This method eliminates prior known information and avoids empirical tuning of the regularization parameters. It involves two major steps: the first step is an unsupervised partitioning using affinity propagation to blindly select spectral components and instead of estimating the PSF using all spectral components, we use only the exemplar component of each correlated group. The second step involves restoration of the hyperspectral image using the appropriate estimated PSF for each group. A multi-scale pyramidal model is used to estimate the PSF and the latent image on each selected spectral component. Hyperspectral images are used to assess the proposed method. Three objective criteria are adopted the mean of PSNR, SSIM, and L1-norm of the estimation error.

Data stream partitioning is an important technique in data mining to analyze data streams in real-time. In this context, lots of data stream partitioning methods have been proposed. In the state of the art, most existing methods need to specify the number of classes before partitioning and/or introduce user-defined parameters for which the parameter values may differ for different data sets. In practice, it is difficult to determine the number of classes and the parameter values. Therefore, we propose in this paper an unsupervised and non-parametric method based on the Optimized Fuzzy C-Means algorithm. It has 2 steps. First is to partition a series of data chunks and then partition the intermediate classes formed before. The performance of the proposed algorithm is evaluated and compared with the recent state-of-the-art methods on hyperspectral image data sets.

In this paper, we propose a new unsupervised and automatic method for the selection of training samples. Thanks to this completely unsupervised method, the samples to be used in the learning task are selected according to objective criteria. Using biased or simplified training samples does not allow a rigorous explanation of the physical phenomena represented by the acquired data, especially in hyperspectral imaging. Furthermore, the use of training samples in learning task is of great importance and essential because they strongly affect the obtained results of any algorithm, when they are simplified or biased. The proposed method was tested on the public IRIS database and on synthetic and real hyperspectral images. Results show that the proposed method can not only select the training samples but also correct the biased or simplified ground truth.

Multichannel systems of remote sensing provide a huge amount of data useful for different applications. However, such images occupy a large space that poses problems of processing, storage, transmission, and management. Lossy compression is widely used to decrease the size of data. In lossy compression, one has to provide a reasonable trade-off between compression ratio (CR) and introduced losses or quality of compressed data. Quality can be characterized in various ways including traditional criteria as peak signal-to-noise ratio (PSNR) or some others as well as criteria that describe efficiency of solving the final tasks of remote sensing as, e.g., probability of correct classification. In this paper, we concentrate on classification of three-channel images that can be either color images or three components of multi- or hyperspectral data acquired, e.g., by Sentinel-2 sensor. In lossy compression of color images, downscaling of color components is often applied to increase CR without essential loss of quality. The goal of this paper is to study the influence of such downscaling on classification accuracy for three-channel remote sensing data. The compression method based on atomic functions is considered since this method allows easy control of compressed image quality and its providing. The neural networks trained for distorted-free images are applied for image classification. Analysis is carried out for four images of different complexity. Based on it, practical recommendations are given.

The exhaustive search of multiple matches in an overlapping area of two multimodal remote sensing images and the accurate localization of found matches are inherent steps to an efficient registration of these two images. A supervised approach based on convolutional neural networks can address this challenge by producing a similarity map, identifying potential matches within a preset search area and estimating a covariation matrix of their location errors. The training is based on a specially designed loss function to enforce the translational and rotational invariance of the similarity map. Using synthetic samples from anisotropic fractal Brownian motion (afBm) models of different orientation, we made the experimental finding that this loss function is biased with respect to orientation. This bias problem is then addressed by beneficially adding pure afBm data to the learning process.

Data stream partitioning has attracted more and more attention in processing large-scale data. The use of parametric methods to perform this task requires for each application an empirical tuning of the different parameters. In practice, this step is difficult to perform and often does not lead to an optimized partitioning result. To avoid this difficulty and provide an objective and optimized partitioning, we propose in this paper an unsupervised and non-parametric algorithm of data stream which employs the Optimized Fuzzy C-Means algorithm. The partitioning is first performed on a series of data chunks and the final partition is obtained by a fusion process of the intermediate classes formed before. The performance of the proposed algorithm is evaluated and compared with a recent state-of-the-art algorithm on hyperspectral image databases.

This paper considers the problem of learning efficient similarity measure (SM) for multimodal remote sensing (RS) images. It is desirable to have a single SM that is efficient for different combinations of modes. We first consider the influence of training dataset balancing on SM efficiency. We demonstrate that it is possible to improve overall SM performance. However, this improvement is observed only for some combinations of modes. To cope with this problem, we propose to include information about the modes compared as additional input to the proposed Convolutional Neural Network (CNN). With this additional information, SM performance for all combinations of modes was improved. We confirm SM efficiency improvement for the real data from Sentinel 2, Landsat 8, Hyperion, SIR-C, and Sentinel 1 platforms, ASTER Global DEM 2, and ALOS World 30m global DEMs and for combinations of modes including optical-to-optical, optical-to-radar, optical-to-DEM and radar-to-DEM and compare the proposed CNN performance with existing SMs.

Many similarity measures (SMs) were proposed to measure the similarity between multimodal remote sensing (RS) images. Each SM is efficient to a different degree in different registration cases (we consider visible-to-infrared, visible-to-radar, visible-to-digital elevation model (DEM), and radar-to-DEM ones), but no SM was shown to outperform all other SMs in all cases. In this article, we investigate the possibility of deriving a more powerful SM by combining two or more existing SMs. This combined SM relies on a binary linear support vector machine (SVM) classifier trained using real RS images. In the general registration case, we order SMs according to their impact on the combined SM performance. The three most important SMs include two structural SMs based on modality independent neighborhood descriptor (MIND) and scale-invariant feature transform-octave (SIFT-OCT) descriptors and one area-based logarithmic likelihood ratio (logLR) SM: the former ones are more robust to structural changes of image intensity between registered modes, the latter one is to image noise. Importantly, we demonstrate that a single combined SM can be applied in the general case as well as in each particular considered registration case. As compared to existing multimodal SMs, the proposed combined SM [based on five existing SMs, namely, MIND, logLR, SIFT-OCT, phase correlation (PC), histogram of orientated phase congruency (HOPC)] increases the area under the curve (AUC) by from 1% to 21%. From a practical point of view, we demonstrate that complex multimodal image pairs can be successfully registered with the proposed combined SM, while existing single SMs fail to detect enough correspondences for registration. Our results demonstrate that MIND, SIFT, and logLR SMs capture essential aspects of the similarity between RS modes, and their properties are complementary for designing a new more efficient multimodal SM.

We consider the problem of density-based unsupervised classification in hyperspectral data. Our focus is especially on methods based on K nearest neighbors (KNN) graph. In this paper, we propose some improvements of recently published methods in this vein, namely GWENN (Graph WatershEd using Nearest Neighbors) as well as a KNN version of Density Peaks Clustering. These improvements address (i) the structure of the KNN graph, which can be modified efficiently to emphasize the dependencies between objects, especially in high dimensional data sets; (ii) the choice of the pointwise density model; and (iii) the ability of these methods to handle variable NN graphs. The improved methods are compared in the context of pixel partitioning in hyperspectral images and are shown to give encouraging results, outperforming state-of-the-art methods like DBSCAN and FCM.

In this article, we borrow from the blind noise parameter estimation (BNPE) methodology early developed in the image processing field an original and innovative no-reference approach to estimate digital elevation model (DEM) vertical error parameters without resorting to a reference DEM. The challenges associated with the proposed approach related to the physical nature of the error and its multifactor structure in DEM are discussed in detail. A suitable multivariate method is then developed for estimating the error in gridded DEM. It is built on a recently proposed vectorial BNPE method for estimating spatially correlated noise using noise informative areas and fractal Brownian motion. The new multivariate method is derived to estimate the effect of the stacking procedure and that of the epipolar line error on local (fine-scale) standard deviation and autocorrelation function width of photogrammetric DEM measurement error. Applying the new estimator to Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) GDEM2 and Advanced Land Observing Satellite (ALOS) World 3D DEMs, good agreement of derived estimates with results available in the literature is evidenced. Adopted for TanDEM-X-DEM, estimates obtained agree well with the values provided in the height error map. In future works, the proposed no-reference method for analyzing DEM error can be extended to a larger number of predictors for accounting for other factors influencing remote sensing (RS) DEM accuracy.

This paper deals with the problem of measuring similarity between visual and radar remote sensing images. It is proposed to combine the benefits of a finite set of representative Similarity Measures (SM) to obtain a combined SM with improved performance in terms of usual assessment criteria (ROC, AUC and LR+). This combined SM relies on a binary linear support vector machines (SVM) classifier trained using real visual-to-radar image pairs RS images. The best combination of SMs among those considered in the finite set is found to be SIFT-OCT, MIND and 10gLR SMs. It reaches a value of AUC criterion about 0.05 higher than that obtained by the best individual SM. This obtained gain is mainly attributed to the complementary properties of structural (SIFT-OCT, MIND) and area-based (logLR) SMs.

Hyperspectral images acquired by remote sensing systems are generally degraded by noise and can be sometimes more severely degraded by blur. In this study, we address the semiblind restoration of the degraded images component-wise, according to a sequential scheme. We propose a new component-wise semi-blind method for estimating effectively and accurately both the blur and the corresponding latent image. To prove applicability and higher efficiency of the proposed method, we compare it against the method it originates from. Our attention is mainly paid to the objective analysis (via l1-norm) of the estimation error accuracy. The tests are performed on a synthetic hyperspectral image. This image has been successively degraded with eight real blurs taken from the literature, each of a different support size. Conclusions, practical recommendations and perspectives are drawn from the results experimentally obtained.

In this communication, we propose a new unsupervised clustering method, which uses a kNN graph to propagate labels, starting from high density regions of the representation space. A feature of this method is the fact that it only requires setting the number of neighbors of each object, a problem which can be addressed easily thanks to the clustering stability of the proposed approach. A multiresolution setting is also proposed to allow clustering image pixels. Preliminary results obtained on real hyperspectral images show the efficiency of the proposed clustering method with respect to classical approaches often used in remote sensing.

This paper deals with prediction of compression ratio (CR) in lossy compression of noisy remote sensing images using techniques based on discrete cosine transform (DCT). Properties of noise assumed additive (in original data or after proper variance stabilizing transform) are taken into account by setting quantization step (QS) proportional to noise standard deviation. It is shown that simple statistics of DCT coefficients in 8×8 blocks can be used for rather accurate prediction of CR. Functions employed in prediction are obtained in advance using curve regression into scatter-plots. The factors that have impact on prediction accuracy are studied. It is demonstrated that percentage of DCT coefficients that become zeroes after quantization can be a good input parameter for prediction. Applicability of the proposed CR prediction approach is confirmed by experiments with real-life multi- and hyperspectral data.

This paper deals with area-based subpixel image registration under the rotation-isometric scaling-translation transformation hypothesis. Our approach is based on parametrical modeling of geometrically transformed textural image fragments and maximum-likelihood estimation of the transformation vector between them. Due to the parametrical approach based on the fractional Brownian motion modeling of the local fragments' texture, the proposed estimator MLfBm(ML stands for “maximum likelihood” and fBm stands for “fractal Brownian motion”) has the ability to better adapt to real image texture content compared with other methods relying on universal similarity measures such as mutual information or normalized correlation. The main benefits are observed when assumptions underlying the fBm model are fully satisfied, e.g., for isotropic normally distributed textures with stationary increments. Experiments on both simulated and real images and for high and weak correlations between registered images show that the MLfBmestimator offers significant improvement compared with other state-of-the-art methods. It reduces translation vector, rotation angle, and scaling factor estimation errors by a factor of about 1.75-2, and it decreases the probability of false match by up to five times. In addition, an accurate confidence interval for MLfBmestimates can be obtained from the Cramér-Rao lower bound on rotation-scaling-translation parameter estimation error. This bound depends on texture roughness, noise level in reference and template images, correlation between these images, and geometrical transformation parameters.

This paper focuses on the potential accuracy of remote sensing (RS) image registration. We investigate how this accuracy can be estimated without ground truth available and used to improve registration quality of mono- and multimodal pair of images. At the local scale of image fragments, the Cramér-Rao lower bound (CRLB) on registration error is estimated for each local correspondence between coarsely registered pair of images. This CRLB is defined by local image texture and noise properties. Opposite to the standard approach, where registration accuracy is only evaluated at the output of the registration process, such valuable information is used by us as an additional input knowledge. It greatly helps in detecting and discarding outliers and refining the estimation of geometrical transformation model parameters. Based on these ideas, a new area-based registration method called registration with accuracy estimation (RAE) is proposed. In addition to its ability to automatically register very complex multimodal image pairs with high accuracy, the RAE method is able to provide registration accuracy at the global scale as a covariance matrix of estimation error of geometrical transformation model parameters or as pointwise registration standard deviation. This accuracy does not depend on any ground truth availability and characterizes each pair of registered images individually. Thus, the RAE method can identify image areas for which a predefined registration accuracy is guaranteed. This is essential for RS applications imposing strict constraints on registration accuracy such as change detection, image fusion, and disaster management. The RAE method is proved successful with reaching subpixel accuracy while registering eight complex mono-/multimodal and multitemporal image pairs including optical-to-optical, optical-to-radar, optical-to-digital elevation model (DEM) images, and DEM-to-radar cases. Other methods employed in comparisons fail to provide in a stable manner accurate results on the same test cases.

We focus on considering noise properties in hyperspectral images acquired by different sensors. An initial assumption is that signal-dependent and signal-independent components are present. Using modern methods of blind estimation of noise parameters from images at hand, contributions of signal-dependent and signal-independent noise components are evaluated and compared for real-life images. It is demonstrated that for some sub-bands, contribution of signal-independent components is prevailing whilst for other sub-band images, the situation is the opposite.

Our paper addresses a question of prediction compression ratio in lossy compression of remote sensing images by coders based on discrete cosine transform (DCT) taking into account noise present in these images. Quantization step is set fixed and proportional to noise standard deviation to provide compression in optimal operation point if it exists. Simple statistics of DCT coefficients is used for predicting compression ratio. Prediction dependences are obtained offline (in advance) and they occur to be quite simple and accurate. The influence of DCT statistics on prediction efficiency is analyzed. Accuracy of prediction is studied for real-life hyperspectral data compressed component-wise.

This paper describes a recently created image database, TID2013, intended for evaluation of full-reference visual quality assessment metrics. With respect to TID2008, the new database contains a larger number (3000) of test images obtained from 25 reference images, 24 types of distortions for each reference image, and 5 levels for each type of distortion. Motivations for introducing 7 new types of distortions and one additional level of distortions are given; examples of distorted images are presented. Mean opinion scores (MOS) for the new database have been collected by performing 985 subjective experiments with volunteers (observers) from five countries (Finland, France, Italy, Ukraine, and USA). The availability of MOS allows the use of the designed database as a fundamental tool for assessing the effectiveness of visual quality. Furthermore, existing visual quality metrics have been tested with the proposed database and the collected results have been analyzed using rank order correlation coefficients between MOS and considered metrics. These correlation indices have been obtained both considering the full set of distorted images and specific image subsets, for highlighting advantages and drawbacks of existing, state of the art, quality metrics. Approaches to thorough performance analysis for a given metric are presented to detect practical situations or distortion types for which this metric is not adequate enough to human perception. The created image database and the collected MOS values are freely available for downloading and utilization for scientific purposes.

International audience

A new performance bound is proposed for analyzing parametric image registration methods objectively. This original bound is derived from the Cramer-Rao lower bound on the estimation error of parameters involved in a geometric transformation assumed between reference and template images (pure translation in this work) and parameters describing the texture of these images. For describing local fragments of both the reference and the template images, the parametric fractional Brownian motion (fBm) model has been chosen. Experimental results, obtained first on pure fBm data with full matching of the data to the texture model assumption, give evidence that the proposed bound describes more adequately the performance of conventional estimators than two other bounds previously proposed in the literature. This holds with respect to the signal-to-noise ratio value of both images, the roughness of their texture, their correlation, and the actual value of translation parameters between their grids. Then, one real Hyperion hyperspectral data set is considered to test the proposed bound behavior on real data. The proposed bound is demonstrated to describe more adequately the estimation accuracy of the translation parameters between different bands of this data set.

The problem of automatic detection of image areas that can be reliably selected for accurate estimation of additive noise standard deviation (STD), irrespectively to processed image properties, is considered in this paper. For getting accurate estimate of either texture or noise parameters involved, we distinguish two complementary image informative maps: (1) noise-informative (NI) map and (2) its complementary texture-informative (TI) map. The NI map is determined and iteratively upgraded based on the Fisher information on noise STD calculated in a single scanning window (SW). The TI map is simply evolved as the complementary part of N map currently updated. Final noise STD estimation is performed by efficient analysis of finite size 9×9 block DCT coefficients in NI SWs. Experiments on large image database have proved that the proposed approach outperforms state-of the-art estimators with respect to both noise STD estimates bias and variance.

We address the problem of unsupervised band reduction in hyperspectral remote sensing imagery. We propose the use of an information theoretic criterion to automatically separate the sensor's spectral range into disjoint subbands without ground truth knowledge. Our approach, named BandClust, preserves the physical sense of the spectral data and automatically provides relevant spectral subbands, i.e., of maximal informational complementarity. Experiments using real hyperspectral images are conducted to compare BandClust with four other unsupervised approaches. The comparison of the selected dimensionality reduction methods is performed via supervised classification using support vector machines and shows the potential of the proposed approach.

The paper addresses blind evaluation of the parameters of mixed noise in images. The conventional approach is based on line fitting in the scatter-plot of local variance estimates using LMS algorithm. This does not utilize the fact that the points in the scatter pot typically appear in clusters that depend on the image. It is shown that the use of weighted LMS algorithm that takes into account the number of points in clusters provides considerable improvement in the accuracy of line fitting and, thus, better estimation of the parameters of mixed noise.

Image segmentation is an important task in image analysis and processing. Many of the existing methods for segmenting a multi-component image (satellite or aerial) are very slow and require a priori knowledge of the image that could be difficult to obtain. Furthermore, the success of each of these methods depends on several factors, such as the characteristics of the acquired image, resolution limitations, intensity in-homogeneities and the percentage of imperfections induced by the process of image acquisition. Recently, fuzzy C-means (FCM) and Genetic Algorithms were separately used in segmenting multi-component images but neither of them had successfully addressed the above concerns. GA was enhanced using Hill-climbing, randomising, and modified mutation operators, leading to what is called hybrid dynamic genetic algorithm (HDGA). Coupling HDGA and FCM creates an unsupervised segmentation method which could successfully segment two types of multi-component images (Landsat ETM+, and IKONOS II). Comparison with the four different methods FCM, hybrid genetic algorithm (HGA), self-organizing-maps (SOM), and the combination of SOM and HGA (SOM-HGA) reveals that FCM-HDGA segmentation method gives robust and reliable results, and is more time efficient.

This communication addresses the problem of the automatic mosaicking of raw images acquired by airborne pushbroom imagers. Using appropriate ancillary data issued from GPS and inertial measurements, we show how the mutual information criterion can be used to improve the co-registration and direct georeferencing of overlapping flight lines by estimating unknown or inaccurate elevation data. The proposed approach does not require any control point to work, and requires only few iterations to improve the initial pose. The mosaicking task itself is performed in a very simple manner. We describe the proposed system, and assess the robustness of our method with an example of application to the mosaicking of multi-track real multispectral image data.

We describe a methodology that is used for the automatic georeferencing of multispectral images acquired from airborne pushbroom imaging cameras. This methodology is based upon the dense registration of flight strips data onto a reference orthoimage and uses the mutual information criterion to make the raw (source) image and the orthoimage (target) aligned. By taking into account a rigorous modeling of the pushbroom imaging process, we show how the mutual information between the raw image and the orthoimage can be used to estimate unknown or inaccurate flight attitude parameters [(tilt bias angles, instantaneous yaw, and height above ground level (AGL)] without using any ground control nor tie points. Moreover, we show how a coarse digital elevation model can be incorporated in the proposed procedure to improve the geocorrection and even estimate a digital surface model of the surveyed area. Two experiments using Compact Airborne Spectrographic Imager (CASI) and Airborne Imaging Spectroradiometer for Applications (AISA) Eagle sensor data are described and show the effectiveness of our approach, with planimetric root mean square errors being in the range of two to four pixels, depending on the flatness and land cover, and altimetric root mean square errors being less than 1% of the flight altitude AGL.

A novel approach to joint estimation of multiplicative noise variance and probability of impulsive noise occurrence in images is proposed. It uses a fractal Brownian motion model for description of real life images. It is demonstrated that this approach provides accurate estimation of mixed noise parameters even for images containing a large percentage of texture regions. The proposed method performance is compared to a modification of a recently designed method based on minimal inter-quantile distances.

Image segmentation is an essential process for image analysis. Several methods were developed to segment multicomponent images, and the success of these methods depends on several factors including (1) the characteristics of the acquired image and (2) the percentage of imperfections in the process of image acquisition. The majority of these methods requirea prioriknowledge, which is difficult to obtain. Furthermore, they assume the existence of models that can estimate its parameters and fit to the given data. However, such a parametric approach is not robust, and its performance is severely affected by the correctness of the utilized parametric model. In this letter, a new multicomponent image segmentation method is developed using a nonparametric unsupervised artificial neural network called Kohonen's self-organizing map (SOM) and hybrid genetic algorithm (HGA). SOM is used to detect the main features that are present in the image; then, HGA is used to cluster the image into homogeneous regions without anya prioriknowledge. Experiments that are performed on different satellite images confirm the efficiency and robustness of the SOM-HGA method compared to the Iterative Self-Organizing DATA analysis technique (ISODATA).

This communication addresses the problem of automatic registration of raw images issued from an airborne multispectral imaging scanner with the aid of reference ortho-images available by classical aerial photography. We develop the proposed model, which accounts for a scan line shift process, in order to compensate for the roll motion of the aircraft, in addition to a RST deformation. The estimation of the model parameters is performed by using a PDE-based approach for the maximization of the mutual information between the source and the target image, starting from an initial estimate of the scan line process. We assess the robustness of this method and show an example of application to aerial image data

Some peculiarities of modified vector sigma filter are studied. In particular, its edge preservation ability is considered in case of processing multichannel remote sensing (RS) images. Such a problem is of high importance for many scene recognition and segmentation tasks. It is demonstrated through comparative quantitative and visual processing data that the proposed filter simultaneously provides efficient noise suppression and excellent edge preservation. Edge detection results that prove this fact are also depicted

This paper deals with the problem of identifying the nature of the noise and estimating its statistical parameters from the observed image in order to be able to apply the most appropriate processing or analysis algorithm afterwards. We focus our attention on three main classes of degraded images, the first one being degraded by an additive noise, the second one by a multiplicative noise, and the latter by an impulse noise. To improve the identification rate, we propose an unsupervised variational classification through a multithresholding method. Each class is then characterized by statistical parameters obtained from homogeneous regions. For the accuracy of the estimation of the noise statistical parameters, we distinguish the corresponding local estimates statistical series according to the number of pixels taken into account to calculate them. The experimental study highlights the improvement so obtained and shows the efficiency and the robustness of the whole method

In this communication, we introduce the concept of gravitational transform, with application to multicomponent image classification. This general concept, from which many settings can be implemented, takes into account both feature sets extracted from the image and the spatial distance between pixels in order to further improve the classification, and thus the partitioning into homogeneous regions. Examples of classification and comparative results with synthetic data are presented, showing that this approach improves the classification rates in all cases, independently of the chosen classification technique, and outperforms some recent approaches in contextual multidimensional unsupervised clustering. The encouraging results obtained make this technique a valuable tool to insert between the feature extraction process and the unsupervised classification process.

We present a new adaptive filtering method for either additive noise or multiplicative noise. The proposed method is stated with a differential equation of the temporal evolution of the problem of interest. It achieves an improvement in the efficiency of the well-known iterated Lee filter in the spatial domain. Mainly, it incorporates a local determination of the optimal regions which are subsequently used to estimate the different local statistics involved in the filtering method. This estimation is carried out differently according to the nature of the processed pixel. An adapted decisional criterion indicates if the pixel belongs either to a contour or to a homogeneous zone, following the idea used in classical anisotropic methods. Then, the efficiency of the proposed method, for which we can prove the existence and uniqueness of the solution, is assessed on several images degraded artificially. The results are compared to the main used filters in order to confirm the theoretical findings.

In this communication, we propose a new approach which enables to fusion either the results of several segmentation methods of a same image or the different results in the case of a multi-components image. The developed method is based on a genetic algorithm approach which allows to combine segmentation results by taking into account their quality through an evaluation criterion. This criterion provides to quantify a segmentation result without any a priori knowledge such as the ground truth. This approach is applied to segment multi-components images by combining the segmentation results of each component. We show the efficiency of the method through some experimental results on several images.

In the domain of multi-bands image processing, two different approaches can be considered: the scalar one and the vectorial one. This paper presents a method that belongs to the first approach. The method is achieved in three steps. The first step tempts to eliminate redundant observations by making a selection of relevant bands. In the second step, each of the selected bands is segmented using a technique of histogram multi-thresholding. In the last step, a fusion by a combination of the results of the selected bands allows one to obtain the final segmentation. This scheme is illustrated in the frame of an application in high-resolution multispectral imagery acquired by the Compact Airborne Spectrographic Imager (CASI).

This paper presents a totally blind restoration system of blurred and noisy images. It includes two processing modules, one for identifying the blur PSF and another for blind image restoration. The blur PSF identification and estimation is achieved using the generalized cross-validation criterion and a smoothness constraint on the image. The deblurring step is based on a Tikhonov-Miller regularization technique. Simulation results show that this system yields rather good results on various kind of images with low signal-to-noise ratio. It is quite effective in restoring severely blurred and noise corrupted images, without prior knowledge of either the noise or image characteristics. It can also discriminate between blurred and unblurred images.

We propose in this communication an unsupervised clustering method called MLBG based upon the K-means algorithm. The originality of this method lies in the automatic determination of the number of clusters by calling into question an intermediate result. This method also enables to improve the different steps in the K-means algorithm. We show the efficiency of the MLBG method through some experimental results and we demonstrate the usefulness of the technique for image segmentation.

In pattern recognition problems, the effectiveness of the analysis depends heavily on the quality of the image to be processed. This image may be blurred and/or noisy and the goal of digital image restoration is to find an estimate of the original image. A fundamental issue in this process is the blur estimation. When the blur is not readily available, it has to be estimated from the observed image. Two main approaches can be found in the literature. The first one identify the blur parameters before any restoration whereas the second one realizes these two steps jointly. We present a comparative study of several parametric blur estimation methods, based on a parametric ARMA modeling of the image, belonging to the first approach. Our purpose is to evaluate the accuracy of the various methods, on which the restoration procedure relies, and their robustness to modeling assumptions, noise, and size of support.

Texture analysis is an important problem in image processing because it conditions the quality of image segmentation and interpretation. We propose in this communication a texture model which is invariant by rotation and whose parameters allow to characterize at the same time the type of texture and its tonal primitive. The originality of the model proposed lies in the use of the Wold decomposition to model the 1D normalized autocovariance. This function is computed from the 2D normalized autocovariance of a texture. Finally, parameters of the model are estimated by using a genetic algorithm. Experimental results on textures from the Brodatz album and synthetic textures show a modeling error lower than 0.06.

This paper deals with the problem of identifying the nature of noise and estimating its standard deviation from the observed image in order to be able to apply the most appropriate processing or analysis algorithm afterwards. In this study, we focus our attention on three classes of degraded noise images, the first one being degraded by an additive noise, the second one by a multiplicative noise and the latter by an impulsive noise. First, in order to identify the nature of the noise, we propose a new approach consisting of characterizing each class by a parameter obtained from histograms computed on several homogeneous regions of the observed image. The homogeneous regions are obtained by segmenting images. Then, the estimation of the standard deviation is achieved from the analysis of an histogram of local standard deviations computed on each of the homogeneous regions.