This paper investigates adaptive coherent detectors in compound Gaussian clutter with Nakagami-distributed textures. The maximum a posterior ratio test (MAPRT) detector is firstly derived. Using the two-step method, the MAP-Rao/Wald detectors are derived. The three detectors have analytical expressions owing to the root formula of depressed cubic equation. The three detectors are proved to be constant false alarm rate (CFAR) with respect to the scale parameter of the texture, speckle covariance matrix, and Doppler steering vector. Their adaptive versions are shown by simulation to be approximate CFAR with estimated speckle covariance matrix by the iterative maximum likelihood (IML) estimator. The performances of the detectors are compared by using simulated and measured data. The results show that the adaptive MAP-Rao and MAPRT detectors obtains better whole performance than MAP-Wald detector and classic adaptive matched filter (AMF) and adaptive normalized matched filter (ANMF) detectors.

For maritime radars to detect low-velocity small targets, it is always difficult to exclude false alarms from sea spikes in observation time of the order of subsecond. This article investigates the characteristics of sea spikes in full-polarimetric maritime radars and proposes a sea-spike discrimination method based on polarimetric Doppler offset differences (DoDs). As the first contribution, an improved sea spike identification method is presented. In the method, the generalized Pareto intensity distribution (GPID) is used to model Bragg scattering background at each polarization and GPID-dependent amplitude threshold, the minimum time width, and range-time connected support are combined to extract sea spikes with range-time supports. As the second contribution, the deep analysis of sea spikes in 70 full-polarimetric IPIX datasets shows that sea spikes synchronously occur at HH-, HV-, and VV-polarizations as a large-scale scattering phenomenon from breaking waves and sea spikes has different Doppler offsets at the three polarizations. Two new indexes of the polarimetric repeatability rate and amplitude peak-mean ratio (APMR) and the concept of full-polarimetric sea spikes are introduced to analyze polarimetric characteristics of sea spikes. As the third contribution, a method of detection of low-velocity small targets and discrimination of sea spikes based on polarimetric DoDs is developed for full-polarimetric maritime radars. It realizes effective detection of low-velocity small targets and exclusion of false alarms from sea spikes in a short coherent processing interval (CPI) of about one-tenth of 1 s.

In this letter, an outlier-robust estimation method using the generalized regression neural network (GRNN) and multiple sample percentiles and truncated moments is proposed to estimate the three parameters of the compound-Gaussian clutter model with generalized inverse Gaussian (CG-GIG) textures. The tri-parametric CG-GIG distributions include K-distributions, generalized Pareto (GP) intensity distributions, and CGIG distributions as biparametric subclasses and thus possess a strong ability to model various sea clutter data without type selection. It is shown that the ratios of the sample percentiles and truncated moments can uniquely be determined by the CG-GIG distribution, despite its overparameterization. The GRNN is trained to generate an approximate inverse function from the ratio vector to the CG-GIG distribution. By the trained GRNN, the GRNN-based outlier-robust parameter estimators are constructed. Experimental results using simulation data and measured sea clutter data show that the GRNN-based estimators outperform the moment-based estimators and the parametric curve fitting estimation (PCFE) method in performance.

Sea clutter modelling and parameter estimation/prediction are important basis of system design, performance assessment, and target detection of maritime radars. Besides parameter estimation from measured data, parameter prediction from radar parameters, sea state, and the viewing geometry of the radar is an alternative way. In this letter, a dual-polarimetric full-recorded sea clutter database measured by an island-based X-band experimental radar on the Yellow sea of China is introduced. On this database, the model selection reveals that the compound-Gaussian model with inverse Gaussian (CGIG) texture is suited for HH high-resolution sea clutter data, the generalized Pareto (GP) intensity distributions are for HH moderate-resolution data, and the K amplitude distributions are for high and moderate-resolution VV polarized data. Further, three five-parametric empirical formulae are constructed to predict the shape parameter of the CGIG, GP, and K distributions of sea clutter from the area of spatial resolution cell, the grazing angle, the significant wave height, the average wave period, and the wave direction relative to the sight line of radar. The new empirical formulae attain more accurate prediction than existing empirical formula on the database.

Sea surface target detection with Graph Neural Networks (GNN) is an emerging method. However, the correlation information of radar returns cannot be efficiently exploited by the conventional Graph Convolutional Network (GCN). Therefore, this paper proposes a small floating target detection method based on graph attention network (GAT) with spatio-temporal correlation of clutter maps, and designs fusion weighting and dynamic attention mechanism for practical problem. First, the dwell radar data is modeled as a graph structure according to its spatio-temporal information. The proposed graph structure allows Doppler spectra of same-type samples to be accumulated separately for sea clutter and target returns. Then, we propose a GAT-based detector and optimize it to create variants: GATv2, GAT-Fused, and GATv2-Fused. These variants aim to reduce sea spike interference and jointly utilize spatio-temporal clutter map information and feature correlation. Both measured and simulated data demonstrate that the proposed attention-based detectors effectively identify marine small floating targets, including out-of-distribution (OOD) detection, outperforming conventional feature-based detectors, the GCN detector, and the pure GAT detector.

This article presents a classification method to classify different marine floating small targets, which can realize effective classification of different targets in strong clutter background. The design of proposed classification method is primarily based on block-whitened time–frequency spectrogram and pre-trained convolution neural network (CNN). Block-whitening clutter suppression is used to process target echoes. By converting a strong clutter background to an approximately noisy background, the effect of strong clutter on classification is reduced. Then, the time–frequency spectrogram of targets is extracted from the block-whitened target echoes, which converts a signal in time domain into a time–frequency spectrogram with more information. In addition, the block-whitened time–frequency spectrograms are input to a pre-trained CNN for feature extraction and classification training. By exploiting pre-training procedure, the proposed method can effectively classify different marine floating small targets and solve the problem of limited target samples in practical applications. Finally, a dataset of three kinds of measured maritime radar targets is constructed to verify the effectiveness of proposed method. Experimental results show that compared with competitors, the pre-trained CNN with block-whitened time–frequency spectrograms can achieve higher performance on the measured dataset.

As a basic characteristic of sea clutter, the reflectivity of sea surface depends on many factors. Various universal empirical models of low precision have been developed to predict the reflectivity of sea surface. In this letter, a method is proposed to train specific predictors by big data learning, where the universal empirical models are embedded to the architecture of the generalized regression neural network (GRNN) to enhance the learning ability and efficiency. On the sea clutter database measured by an island-based UHF-band radar in the offshore waters of the Yellow sea of China at low grazing angle, the GRNN-based predictors of different structures are compared with other predictors. The results on the database show that the GRNN-based predictors behave better at learning efficiency, prediction precision, and robustness.

In this paper, a spatial-temporally varying generalized Pareto intensity distribution (STV-GPID) model is presented to characterize large-scene sea clutter in high-resolution maritime surveillance radars, and a multiscan recursive Bayesian bipercentile (MSRB-BiP) estimation method is proposed to implement the outlier-robust estimation of parameters in the STV-GPID model. Considering that sea clutter characteristics are affected by sea states and the viewing geometry of a radar, the large scene is segmented into clutter map cells based on an empirical backscattering coefficient model of sea surface to predict radar cross-section of sea surface per unit physical area. Sea clutter intensities on each clutter map cell are modelled by a generalized Pareto intensity distribution. In the parameter estimation, the data of previous scans are transformed into the prior information on the parameters to reduce the storage burden of radar systems. The MSRB-BiP estimator updates the parameters of the STV-GPID model recursively by a mixed sample set with the returns of the present scan and simulated data using the prior information. The mixture ratio adjusts the forgetting rate of data to adapt to temporally varying characteristics of sea clutter. At least, it brings three merits: low storage requirement, outlier-robustness, and mitigation of spatial small sample size of a single scan. The convergence and robustness of the estimation method are verified by simulated data. The experimental results on two measured radar datasets verify the effectiveness of the MSRB-BiP estimators and the errors at the steady state are reduced at least 17.7% and 66.7%, respectively.

This letter proposes a region-merging-based method for synthetic aperture radar (SAR) image segmentation, where the merging cost is a fusion of texture pattern similarity measure (TPSM), statistical similarity measure (SSM), and the relative common boundary length penalty (RCBLP). The segmentation is implemented in three steps. First, an image is oversegmented based on the multiscale Bhattacharyya distance to generate an initial partition of considerable regions. Second, regions with sizes under a given threshold are mandatorily merged to yield a middle segmentation. Third, a region-merging process using the new merging cost is iteratively conducted to achieve final segmentation. Due to the existence of the TPSM in the merging cost, the new method avoids the false merging of adjacent regions with different textures. Experimental results of the real SAR images show that the proposed method outperforms existing region-merging-based methods.

Radar external calibration is an important measurement in determining the radar cross sections of unknown targets and the reflectance coefficients of the sea surface. This measurement is often made in an anechoic chamber by standard programs. Sometimes, this measurement must be made in outdoor conditions, and the ambient environment severely affects the calibration precision. Thus, the scheme of the measurement must be specially designed. In this article, a calibration conducted at the sea near Lingshan Island is reported, and a full program of an external calibration on the sea surface and relevant analysis are presented for a P-band island-based sea clutter measurement radar. Data analysis shows that the multipath effect from the sea surface is a major factor to degrade the precision of the calibration. Measurement data over 4 days and the multipath reflectance model of the sea surface are combined to estimate the power increment from the multipath effect. Radar system constant is estimated based on the estimated power increments at different sites, where the viewing geometries of the radar are different. The results show that the proposed external calibration method on the sea surface attains a satisfactory precision.

In this letter, we focus on the statistical modeling of sea clutter amplitudes. Due to its non-Gaussian nature, the existing statistical models are sometimes difficult to represent well the heavy-tailed portion of amplitude distribution. To address this problem, we propose a compound Gaussian (CG) model with a generalized inverse Gaussian (GIG) texture to describe sea clutter amplitudes. In this regard, the probability density function and the cumulative distribution function of the clutter amplitudes for the proposed model are derived. Moreover, we provide an approach to estimate the unknown parameters of the proposed CG-GIG distribution. The experimental results indicate that the CG-GIG distribution is more suitable to describe the amplitudes of non-Gaussian sea clutter than its competitors.

It is a difficult task for high-resolution maritime radars operating at a fast scan mode to find sea-surface floating and low-velocity small targets due to ubiquitous sea clutter, sporadic sea spikes like target returns, and shortage of shared database. In this paper, a simulation method is presented to generate high-resolution radar returns of a local sea surface with a structural trend in texture and sea spikes by integrating the existing results on large-scale sea surface generation, sea surface reflectivity, Doppler characteristics of sea clutter, and properties of sea spikes. Generally, sea-surface small target detection at a fast scan mode is composed of intrascan integration to suppress sea clutter and interscan integration to exclude false alarms and sea spikes. Based on the Doppler difference between targets and sea clutter at the two time scales of a coherent processing interval (CPI) of tens of milliseconds and a scan period of several seconds, a Doppler-guided retrospective filter (DGRF) detector is proposed, which uses the optimum coherent detection in intrascan integration and a DGRF in interscan binary and test statistic integrations. The two integrations and Doppler consistency of integrated plots are fused for the final decision. Owing to the Doppler guidance, the proposed detector effectively impedes the integration of false alarms of the intrascan processing and provides significant detection performance improvement, which is verified by the simulated data and real radar data with test small targets.

In this paper, a novel edge detector for synthetic aperture radar (SAR) images is proposed by introducing the Bhattacharyya coefficient (BC) combining with the rotated biwindow configuration. Based on the quantified input image, the BC is computed from two sample distribution histograms of local regions supported by the subwindows on the opposite sides of the pixel to be detected. With biwindows of different directions sliding through the image, multiple directional Bhattacharyya coefficient matrices are obtained, which are utilized to extract the edge strength map (ESM), characterizing the intensity variation in SAR images. Sequent nonmaximum suppression and hysteresis thresholding refine the extracted ESM into thin edges. Experiment results show that the proposed edge detector can accurately extract edges. Moreover, the BC-based ESM can act as a good precursor to guide SAR image segmentation based on region merging.

This letter addresses the feature detection design for the target embedded in sea clutter. Three polarization features (the relative surface scattering power, the relative volume scattering power, and the relative dihedral scattering power) are obtained based on the observed multipolarization channel returns. Then, 3-D feature detector is constructed to detect the sea-surface small floating target based on the three polarization features. The experiments based on the measured radar data show that the proposed method attains better detection performance and better robustness than do several existed feature-class detectors.

This paper presents one feature-based detector to find sea-surface floating small targets. In integration time of the order of seconds, target returns exhibit time-frequency (TF) characteristics different from sea clutter. The normalized smoothed pseudo-Wigner-Ville distribution (SPWVD) is proposed to enhance TF characteristics of target returns, which is computed from the SPWVDs of time series at the cell under test (CUT) and reference cells around the CUT. The differences between target returns and the TF pattern of sea clutter are congregated on the normalized SPWVD. From that the ridge integration (RI) is computed and significant TF points from each time slice form a binary image. The number of connected regions and the maximum size of connected regions in the binary image are extracted and are combined with the RI into a 3-D feature vector. Due to the unavailability of the feature vector samples of radar returns with target, a one-class classifier with a controllable false alarm rate is constructed from the feature vector samples of sea clutter by the fast convex hull learning algorithm. As a result, a new feature-based detector is designed. It is compared with the tri-feature-based detector using amplitude and Doppler features and the fractal-based detector using the Hurst exponent of amplitude time series on the recognized IPIX radar database for floating small target detection. The results show that a significant improvement in detection performance is attained.

Traditional differential-based edge detection suffers from abrupt degradation in performance when images are corrupted by impulse noises. The morphological operators such as the median filters and weighted median filters possess the intrinsic ability to counteract impulse noise. In this paper, by combining the biwindow configuration with weighted median filters, anisotropic morphological directional derivatives (AMDDs) robust to impulse noise are proposed to measure the local grayscale variation around a pixel. For ideal step edges, the AMDD spatial response and directional representation are derived. The characteristics and edge resolution of two kinds of typical biwindows are analyzed thoroughly. In terms of the AMDD spatial response and directional representation of ideal step edges, the spatial matched filter is used to extract the edge strength map (ESM) from the AMDDs of an image. The spatial and directional matched filters are used to extract the edge direction map (EDM). Embedding the extracted ESM and EDM into the standard route of the differential-based edge detection, an anti-impulse-noise AMDD-based edge detector is constructed. It is compared with the existing state-of-the-art detectors on a recognized image dataset for edge detection evaluation. The results show that it attains competitive performance in noise-free and Gaussian noise cases and the best performance in impulse noise cases.

In this study, the lifting scheme is first employed to design two‐channel biorthogonal graph filter bank. The biorthogonal condition is parameterised by imposing a single‐level lifting structure on the analysis and synthesis graph kernels. Based on the parametric structure, the two kernels are separately optimised by constrained quadratic programming. The obtained two‐channel biorthogonal graph filter banks are of structurally perfect reconstruction. Numerical results and comparison are included to show the proposed algorithm can lead to biorthogonal graph filter banks with improved performance.

A novel non‐coherent detection scheme for range‐spread targets in the sea clutter is developed in this study. The detector utilises the bi‐window non‐linear shrinkage map (BNSM) to reduce sea clutter while preserving target echoes before energy integration (EI). The EI detector based on BNSM (EI‐BNSM) is compared with the conventional non‐coherent constant false alarm ratio detectors based on measured data, and the good performance of EI‐BNSM detector is reported.

We propose a low complexity robust beamforming method for the general-rank signal model, to combat against mismatches of the desired signal array response and the received signal covariance matrix. The proposed beamformer not only considers the norm bounded uncertainties in the desired and received signal covariance matrices, but also includes an additional positive semidefinite constraint on the desired signal covariance matrix. Based on the worst-case performance optimization criterion, a computationally simple closed-form weight vector is obtained. Simulation results verify the validity and robustness of the proposed beamforming method.

Compound-Gaussian model with the inverse Gaussian texture (IG-CG) is recognized to be one of the best models to characterize high-resolution sea clutter at low grazing angles. The model parameters are often estimated by the second- and fourth-order amplitude sample moments, which are of low precision and easily interfered by outliers of high power such as returns of ships and reefs and sea spikes. In this letter, an iterative maximum likelihood (ML) estimator and an outlier-robust bipercentile estimator are proposed and are compared with the moment-based estimator. The experimental results show that the iterative ML estimator is better in performance than the moment-based estimator when samples are without outliers and the bipercentile estimator behaves better when samples contain a small number of outliers.

In this paper, the amplitude time series of sea clutter are described by the extended self-similar (ESS) process and the extended fractal (EF) feature of real-life radar datasets is evaluated. Different from the fractal description of sea clutter by fractal dimension or Hurst parameter at a finite scale range, the EF feature can characterize the roughness of the clutter at different scales by a series of multiscale Hurst parameters. A detail analysis of real-life sea clutter time series is made by multiscale Hurst parameters, which expose the dominating factor at three specific scale ranges. A detector via multiscale Hurst parameter is proposed to detect the floating target in sea clutter and the performance is verified by real-life IPIX radar datasets.

In this study, the authors consider the problem of digital modulation classification in a multipath channel using the cumulant‐based method. A new estimator is proposed to estimate the fourth‐order cumulant of transmitted symbol in a multipath channel with the help of the fourth‐order cross‐cumulants of the received symbols. The proposed estimator does not need the estimation of the multipath channel coefficients and is shown to be asymptotically unbiased. The channel order is an important factor to affect the estimator. The impact of the channel order mismatch on the estimator is analysed and it is proved that the estimator keeps unbiased in the case of the overestimation of the channel order. The simulated experiments are made to verify the cumulant‐based classifier using the new estimator and the results show that it attains better performance for digital modulation classification than the two existing cumulant‐based classifiers in a multipath channel.

Detection of manoeuvring weak targets in radars often encounters circumstance where target movement is modelled by non‐linear dynamic systems and received returns are corrupted by background noise of unknown statistics. It is known that the cost‐reference particle filter (CRPF) is an efficient algorithm for state estimation of non‐linear dynamic systems of unknown statistics. By combining an approximate logarithm likelihood ratio under the piecewise parametric model of signals with the CRPF algorithm, this study proposes a new track‐before‐detect detector, named CRPF‐based detector, for manoeuvring weak target detection from received returns corrupted by background noise of unknown statistics. Experiments using simulated noise and real background noise of over‐the‐horizon radar are made to verify the CRPF‐based detector. The results show that the CRPF‐based detector has comparable performance with the two PF‐based detectors for background noise of known statistics. For background noise of unknown statistics, the CRPF‐based detector attains better detection performance than the two PF‐based detectors where an assumptive probabilistic model is imposed on the background noise.

In this paper, a region-merging-based method is proposed for fast segmentation of amplitude-format synthetic aperture radar (SAR) images. It combines the existing fast initial partition by applying the watershed transform to the thresholded ratio edge strength map with fast region merging by using a new merging cost with relative common boundary length penalty (RCBLP) and the nearest neighbor graph (NNG) for fast search minimal edge on a region adjacency graph (RAG). A new statistical similarity measure, which is a scale-invariant and approximately constant false alarm rate with respect to region sizes, is proposed and combined with an RCBLP term to form a new merging cost. The region-merging process starting from the initial partition is fast implemented by means of the RAG and NNG. Several quantitative indexes in optical image segmentation assessment are borrowed for quantitative assessment of segmentation quality. Experiments to synthetic and real SAR images are reported. The results show that the proposed method is fast and attains higher quality segmentation results than the two recent state-of-the-art methods.

Automatic modulation classification is a key technique in wireless communication systems. In practical scenario, the modulation format of a received signal may be new or unknown to classifiers. However, most of existing classifiers compulsively classify the received signal as one of the candidate modulation formats. The rejection ability of a classifier to unknown modulation formats is important in applications. In this paper, a fourth-order cumulant-based classifier with rejection ability is proposed to classify digital modulation formats under additive white Gaussian noise channel. The two fourth-order cumulants are used as the feature vector. The classification with rejection ability boils down to the segmentation problem of the two-dimensional feature space with rejection region. A two-stage optimization is proposed to attain the suboptimal solution of the problem. The greedy convexhull learning algorithm is used to determine the primary decision regions of all the candidate modulation formats from training sets. The primary decision regions have rejection ability but are not always mutually separate. The alternative convexhull shrinkage is presented to separate the primary decision regions at small loss in probability of correct classification (PCC). The proposed classifier is compared with the cumulant-based hierarchical classifier and the K-S classifiers. The results show that besides desired rejection ability it is competitive with these classifiers in PCC.

This paper proposes a corner detector and classifier using anisotropic directional derivative (ANDD) representations. The ANDD representation at a pixel is a function of the oriented angle and characterizes the local directional grayscale variation around the pixel. The proposed corner detector fuses the ideas of the contour- and intensity-based detection. It consists of three cascaded blocks. First, the edge map of an image is obtained by the Canny detector and from which contours are extracted and patched. Next, the ANDD representation at each pixel on contours is calculated and normalized by its maximal magnitude. The area surrounded by the normalized ANDD representation forms a new corner measure. Finally, the nonmaximum suppression and thresholding are operated on each contour to find corners in terms of the corner measure. Moreover, a corner classifier based on the peak number of the ANDD representation is given. Experiments are made to evaluate the proposed detector and classifier. The proposed detector is competitive with the two recent state-of-the-art corner detectors, the He & Yung detector and CPDA detector, in detection capability and attains higher repeatability under affine transforms. The proposed classifier can discriminate effectively simple corners, Y-type corners, and higher order corners.

By introducing Gaussian-Gamma-shaped (GGS) bi-windows instead of traditional rectangle bi-windows, a new ratio-based edge detector is proposed to extract thin edges of synthetic aperture radar (SAR) images. As poor 2-D smoothing filters, the rectangle window functions are shown to be apt to incur false-edge pixels near true edges. Using the GGS window functions reduces false-edge pixels near true edges, which can be verified by analyzing effective false maxima in the edge strength maps (ESMs). Operating the nonmaximum suppression and hysteresis thresholding on the ratio-based ESM using GGS bi-windows yields thin edges of SAR images. The receiver-operating-characteristic curve is used to evaluate edge detectors. The experimental results to a synthetic SAR image show that the detector using GGS bi-windows attains better performance than the one using rectangle bi-windows.

The direction–frequency selectivity of two-dimensional (2-D) separable oversampled discrete Fourier transform (DFT) modulated filter banks, showing that such 2-D filter banks can provide fine frequency tiling and efficient image representations with direction–frequency selectivity, is analysed. Moreover, the doubly local Wiener filtering is extended to the case using two 2-D separable oversampled DFT modulated filter banks, where the empirical subband energy distributions of the image are estimated by the two sets of directional windows matching the direction–frequency selectivity of the subband filters. The proposed algorithm is of low computational complexity and exhibits a great capability to preserve inhomogeneous textures, owing to the fact that 2-D separable DFT modulated bases are suited to represent oscillating patterns in images. The experimental results show that the proposed denoising algorithm is competitive with the existing algorithms with comparable computational complexity. Particularly, for images with abundant inhomogenous texture, it gives larger output peak signal-to-noise ratios (PSNRs) and achieves better visual effect in texture regions of images.

We propose an analytic search method for interferometric synthetic aperture radar (InSAR) image registration. An analytic cost function is established, and the subpixel offsets associated with the maximum of the cost function are continuously searched by direction-alternate optimization methods along vertical and horizontal directions. We establish a novel dual-quartic analytic cost function for extraction of subpixel offsets from a corresponding pixel pair established by pixel-level registration. A robust optimization method that is called biiteration algorithm for searching the maximum point of the dual-quartic cost function is developed to solve the subpixel offsets. Since the dual-quartic cost function is quartic if one of two parameters is fixed, one of two substeps including in each iterative step of the efficient biiterative method is to find the solution by the secant method. The performances of the proposed method are shown by using simulated and real data and compared with those of representative existing registration method.

In this letter, variational models in pixel domain and wavelet domain are presented for fusion and denoising of noisy multifocus images. In pixel domain, the problem is formulized as minimizing a weighted energy functional, where the total variation (TV) is used as regularity constraint for noise reduction. A new family of weight functions for fusion is proposed that are based on the local average modulus of gradients and the power transform. In wavelet domain, the problem is formulized as shrinkage of the weighted wavelet coefficients of source images, where weight functions are based on the local average modulus of intra- and inter-scale wavelet coefficients and the power transform. The experiments are made to verify the effectiveness of the proposed methods.

The Chan-Vese model is very efficient in segmenting images. However, the algorithm given by Chan and Vese is sensitive to the initial level set function and the regularization parameter. It is difficult to get a right segmentation if the initial level set function and the regularization parameter are not chosen properly. In this paper, we aim to automatically and accurately segment binary images . We propose a two-stage segmentation algorithm and an adaptive parameter estimation method for the regularization parameter. Experiments on some synthetic images and real images show that the proposed algorithm is very efficient.

Local Wiener filtering in the wavelet domain is an effective image denoising method of low complexity. In this letter, we propose a doubly local Wiener filtering algorithm, where the elliptic directional windows are used for different oriented subbands in order to estimate the signal variances of noisy wavelet coefficients, and the two procedures of local Wiener filtering are performed on the noisy image. The experimental results show that the proposed algorithm improves the denoising performance significantly.