Kacem Chehdi

Noise Identification and Estimation of its Statistical Parameters by Using Unsupervised Variational Classification

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

Prediction Of Optimal Operation Point Existence And Parameters In Lossy Compression Of Noisy Images

This paper deals with lossy compression of images corrupted by additive white Gaussian noise. For such images, compression can be characterized by existence of optimal operation point (OOP). In OOP, MSE or other metric derived between compressed and noise-free image might have optimum, i.e., maximal noise removal effect takes place. If OOP exists, then it is reasonable to compress an image in its neighbourhood. If no, more “careful” compression is reasonable. In this paper, we demonstrate that existence of OOP can be predicted based on very simple and fast analysis of discrete cosine transform (DCT) statistics in 8x8 blocks. Moreover, OOP can be predicted not only for conventional metrics as MSE or PSNR but also for visual quality metrics. Such prediction can be useful in automatic compression of multi- and hyperspectral remote sensing images.

VST-based lossy compression of hyperspectral data for new generation sensors

This paper addresses lossy compression of hyperspectral images acquired by sensors of new generation for which signaldependent component of the noise is prevailing compared to the noise-independent component. First, for sub-band (component-wise) compression, it is shown that there can exist an optimal operation point (OOP) for which MSE between compressed and noise-free image is minimal, i.e., maximal noise filtering effect is observed. This OOP can be observed for two approaches to lossy compression where the first one presumes direct application of a coder to original data and the second approach deals with applying direct and inverse variance stabilizing transform (VST). Second, it is demonstrated that the second approach is preferable since it usually provides slightly smaller MSE and slightly larger compression ratio (CR) in OOP. One more advantage of the second approach is that the coder parameter that controls CR can be set fixed for all sub-band images. Moreover, CR can be considerably (approximately twice) increased if sub-band images after VST are grouped and lossy compression is applied to a first sub-band image in a group and to “difference” images obtained for this group. The proposed approach is tested for Hyperion hyperspectral images and shown to provide CR about 15 for data compression in the neighborhood of OOP.

Statistical power of intensity- and feature-based similarity measures for registration of multimodal remote sensing images

This paper investigates performance characteristics of similarity measures (SM) used in image registration domain to discriminate between aligned and not-aligned reference and template image (RI and TI) fragments. The study emphasizes registration of multimodal remote sensing images including optical-to-radar, optical-to-DEM, and radar-to- DEM scenarios. We compare well-known area-based SMs such as Mutual Information, Normalized Correlation Coefficient, Phase Correlation, and feature-based SM using SIFT and SIFT-OCT descriptors. In addition, a new SM called logLR based on log-likelihood ratio test and parametric modeling of a pair of RI and TI fragments by the Fractional Brownian Motion model is proposed. While this new measure is restricted to linear intensity change between RI and TI (assumption somewhat restrictive for multimodal registration), it takes explicitly into account noise properties of RI and TI and multivariate mutual distribution of RI and TI pixels. Unlike other SMs, distribution of logLR measure for the null hypothesis does not depend on registration scenario or fragments size and follows closely chi-squared distribution according to Wilks’s theorem. We demonstrate that a SM utility for image registration purpose can be naturally represented in (True Positive Rate, Positive Likelihood Rate) coordinates. Experiments on real images show that overall the logLR SM outperforms the other SMs in terms of area under the ROC curve, denoted AUC. It also provides the highest Positive Likelihood Rate for True Positive Rate values below 0.4-0.6. But for certain registration problem types, logLR can be second or third best after MI or SIFT SMs.

Scatter-plot-based method for noise characteristics evaluation in remote sensing images using adaptive image clustering procedure

Several modifications of scatter-plot-based method for mixed noise parameters estimation are proposed. The modifications relate to the stage of image segmentation and they are intended to adaptively separate image blocks into clusters taking into account image peculiarities and to choose a required number of clusters. Comparative performance analysis of the proposed modifications for images from TID2008 database is performed. It is shown that the best estimation accuracy is provided by a method with automatic determination of a required number of clusters followed by block separation into clusters using k-means method. This modification allows improving the accuracy of noise characteristics estimation by up to 5% for both signal-independent and signal-dependent noise components in comparison to the basic method. The results for real-life data are presented.

Accuracy assessment of blind and semi-blind restoration methods for hyperspectral images

Hyperspectral images acquired by remote sensing systems are generally degraded by noise and can be sometimes more severely degraded by blur. When no knowledge is available about the degradations present or the original image, blind restoration methods must be considered. Otherwise, when a partial information is needed, semi-blind restoration methods can be considered. Numerous semi-blind and quite advanced methods are available in the literature. So to get better insights and feedback on the applicability and potential efficiency of a representative set of four semi-blind methods recently proposed, we have performed a comparative study of these methods in objective terms of blur filter and original image error estimation accuracy. In particular, we have paid special attention to the accurate recovering in the spectral dimension of original spectral signatures. We have analyzed peculiarities and factors restricting the applicability of these methods. Our tests are performed on a synthetic hyperspectral image, degraded with various synthetic blurs (out-of-focus, gaussian, motion) and with signal independent noise of typical levels such as those encountered in real hyperspectral images. This synthetic image has been built from various samples from classified areas of a real-life hyperspectral image, in order to benefit from realistic reference spectral signatures to recover after synthetic degradation. Conclusions, practical recommendations and perspectives are drawn from the results experimentally obtained.

Potential accuracy of translation estimation between radar and optical images

This paper investigates the potential accuracy achievable for optical to radar image registration by area-based approach. The analysis is carried out mainly based on the Cramér–Rao Lower Bound (CRLB) on translation estimation accuracy previously proposed by the authors and called CRLBfBm. This bound is now modified to take into account radar image speckle noise properties: spatial correlation and signal-dependency. The newly derived theoretical bound is fed with noise and texture parameters estimated for the co-registered pair of optical Landsat 8 and radar SIR-C images. It is found that difficulty of optical to radar image registration stems more from speckle noise influence than from dissimilarity of the considered kinds of images. At finer scales (and higher speckle noise level), probability of finding control fragments (CF) suitable for registration is low (1% or less) but overall number of such fragments is high thanks to image size. Conversely, at the coarse scale, where speckle noise level is reduced, probability of finding CFs suitable for registration can be as high as 40%, but overall number of such CFs is lower. Thus, the study confirms and supports area-based multiresolution approach for optical to radar registration where coarse scales are used for fast registration “lock” and finer scales for reaching higher registration accuracy. The CRLBfBm is found inaccurate for the main scale due to intensive speckle noise influence. For other scales, the validity of the CRLBfBm bound is confirmed by calculating statistical efficiency of area-based registration method based on normalized correlation coefficient (NCC) measure that takes high values of about 25%.

Performance prediction for 3D filtering of multichannel images

Performance of denoising based on discrete cosine transform applied to multichannel remote sensing images corrupted by additive white Gaussian noise is analyzed. Images obtained by satellite Earth Observing-1 (EO-1) mission using hyperspectral imager instrument (Hyperion) that have high input SNR are taken as test images. Denoising performance is characterized by improvement of PSNR. For hard-thresholding 3D DCT-based denoising, simple statistics (probabilities to be less than a certain threshold) are used to predict denoising efficiency using curves fitted into scatterplots. It is shown that the obtained curves (approximations) provide prediction of denoising efficiency with high accuracy. Analysis is carried out for different numbers of channels processed jointly. Universality of prediction for different number of channels is proven.

Efficiency analysis for 3D filtering of multichannel images

Modern remote sensing systems basically acquire images that are multichannel (dual- or multi-polarization, multi- and hyperspectral) where noise, usually with different characteristics, is present in all components. If noise is intensive, it is desirable to remove (suppress) it before applying methods of image classification, interpreting, and information extraction. This can be done using one of two approaches – by component-wise or by vectorial (3D) filtering. The second approach has shown itself to have higher efficiency if there is essential correlation between multichannel image components as this often happens for multichannel remote sensing data of different origin. Within the class of 3D filtering techniques, there are many possibilities and variations. In this paper, we consider filtering based on discrete cosine transform (DCT) and pay attention to two aspects of processing. First, we study in detail what changes in DCT coefficient statistics take place for 3D denoising compared to component-wise processing. Second, we analyze how selection of component images united into 3D data array influences efficiency of filtering and can the observed tendencies be exploited in processing of images with rather large number of channels.

A stable and unsupervised Fuzzy C-Means for data classification

In this paper a stable and unsupervised version of FCM algorithm named FCMO is presented. The originality of the proposed FCMO algorithm relies: i) on the usage of an adaptive incremental technique to initialize the class centres that calls into question the intermediate initializations; this technique renders the algorithm stable and deterministic, and the classification results do not vary from a run to another, and ii) on the unsupervised evaluation criteria of the intermediate classification result to estimate the optimal number of classes; this makes the algorithm unsupervised. The efficiency of this optimized version of FCM is shown through some experimental results for its stability and its correct class number estimation.