Patch Clustering Research Articles

Exploring the spatial patterns of vegetation fragmentation using local spatial autocorrelation indices

The spatial heterogeneity of urban vegetation obtained from discrete land cover classes is sensitive to classification errors and can result in a substantial loss of information, due to the degradation of continuous quantitative information. Although there is an increasing ecological need to use continuous methods to understand spatial heterogeneity and vegetation fragmentation, they remain unexplored. Since local indicators of spatial association (LISA) can capture important spatial patterns of clustering and dispersion at a local scale, it can capture important ecological patterns and process of vegetation fragmentation. This work examines the utility of LISA which allows exploration of local patterns in spatial data in identifying high (hot spots) and low (cold spots) spatial clusters of vegetation patches and fragmentation patterns in Harare metropolitan city in Zimbabwe. The LISA indices of Getis-Ord Gi* and local Moran’s I are computed both on continuous normalized difference vegetation index and discrete land cover data of vegetation and nonvegetation of Sentinel 2016 and 2018. Local spatial clustering patterns are identified with Z-score values that indicate the significance of each statistic. High positive Z-scores are located in the large core, undisturbed, and homogeneous vegetation. Negative Z-scores are located in more dispersed and highly fragmented vegetation. The results suggest that there is a strong tendency for large core, undisturbed, and homogeneous vegetation patches to be spatially clustered and for small, isolated and sparse vegetation patches to be dispersed. The highly fragmented vegetation patches are located in the heavily urbanized part of the city. Overall, findings of this study underscore the relevance of the spatially explicit method of LISA as a valuable source of spatial information for the assessment of local spatial clustering and dispersion of urban vegetation patches. This can be used to develop policies that support effective conservation and restoration strategies.

External prior learning and internal mean sparse coding for image denoising

Image prior and sparse coding learning methods have important uses in image denoising. Many denoising methods learn priors either from the noisy image itself or an external clean image dataset. But using only these as priors does not always reconstruct the image effectively. In addition, when the image is corrupted by noise, the local sparse coding coefficient obtained from a noisy image patch is inaccurate, restricting denoising performance. We present a noise removal framework based on external prior learning and an internal mean sparse coding method, making use of the innate sparsity and nonlocal self-similarity (NSS) of natural images. Specifically, we first obtain external priors from a clean natural image dataset by Gaussian mixture model. The external priors are applied to guide the subspace clustering of internal noisy image patches, and a compact dictionary is generated for each internal noisy patch cluster. Then an internal mean sparse coding strategy based on NSS is introduced into the sparse representation model, whose regularization parameters then are deduced through a Bayesian framework. An iterative shrinkage method is employed to solve the l1-optimization problem in the sparse representation model. Application of the noise removal model to 16 test images demonstrates denoising performance exceeding other competing methods.

Ensemble Dictionary Learning for Single Image Deblurring via Low-Rank Regularization.

Sparse representation is a powerful statistical technique that has been widely utilized in image restoration applications. In this paper, an improved sparse representation model regularized by a low-rank constraint is proposed for single image deblurring. The key motivation for the proposed model lies in the observation that natural images are full of self-repetitive structures and they can be represented by similar patterns. However, as input images contain noise, blur, and other visual artifacts, extracting nonlocal similarities only with patch clustering algorithms is insufficient. In this paper, we first propose an ensemble dictionary learning method to represent different similar patterns. Then, low-rank embedded regularization is directly imposed on inputs to regularize the desired solution space which favors natural and sharp structures. The proposed method can be optimized by alternatively solving nuclear norm minimization and norm minimization problems to achieve higher restoration quality. Experimental comparisons validate the superior results of the proposed method compared with other deblurring algorithms in terms of visual quality and quantitative metrics.

MRI Denoising Using Low Rank Prior and Sparse Gradient Prior

Image priors have been successfully introduced to solve ill-posed problems, such as image denoising. In this paper, we propose a new denoising model for magnetic resonance images (MRIs) which employs the image low-rank and sparse gradient priors. First, we use a Gaussian mixture model (GMM) to guide the clustering of non-local self-similar patches by learning the structure of external noise-free MRI patches to help retain the low rank of the noisy MRI patch matrix. Second, we fit the heavy-tailed gradient of MRI with a hyper-Laplacian distribution to reduce ringing artifacts. Third, we adopt an alternating iterative algorithm. The experimental results show that our proposed algorithm outperforms many classical MRI denoising methods, such as unbiased nonlocal means (UNLM) filtering and block-matching and 3D (BM3D) filtering in both visual and numerical results.

Patch-Based Nonlocal Adaptive Gradient Regularization for Image Restoration in Sensor Networks

In this paper, we propose a novel patch-based adaptive nonlocal gradient regularization method for image restoration in sensor networks. It formulates the hyper-Laplacian distribution to regularize the global gradient distribution. The patch-based nonlocal gradient prior is utilized to regularize the nonlocal self-similarity of image gradients. Firstly, the L0-norm smoothing scheme is used innovatively as the preprocessing step to preserve strong edges, which are critical to improve the accuracy of clustering the similar image patches. Then, adaptive weights for each patches are developed from a set of clustered nonlocal self-similarity patches by learning the the expectation and variance for sparse gradient distribution at each pixel. Comparing with several recent state-of-the-art methods, experimental results show that the proposed method has better performance in alleviating block effects and preserving image details.

Unsupervised Segmentation of Point Clouds From Buildings Using Hierarchical Clustering Based on Gestalt Principles

Segmentation is a fundamental step for parsing the point clouds of three-dimensional (3-D) scenes, which normally contain a wide variety of complex objects and structures with a large amount of points. In this paper, we propose a voxel-based point cloud segmentation method using Gestalt principles under a hierarchical clustering framework, allowing a completely automatic but parametric process for segmenting 3-D scenes of buildings. The voxel-based data structure can increase the efficiency and robustness of the segmentation process. By the use of Gestalt principles, our method conducts the segmentation in a pure geometric way avoiding the use of RGB color and intensity information, which can be applied to general applications. The clustering of patches in our method is carried out on the basis of the local geometric information, which is modeled by the probabilistic formulation and solved by the graphical model. Experiments using terrestrial laser scanning dataset have demonstrated that our proposed method can achieve good results, especially for complex scenes and nonplanar surfaces of objects. The quantitative comparison between our method and other representative segmentation methods (i.e., region growing, voxel-based incremental segmentation, locally convex connected patches, etc.) also confirms the effectiveness and efficiency of our method, with overall $F_1$ measures better than 0.66 for our datasets under complex situations of urban scenes with irregular shaped buildings.

HCLR: A hybrid clustering and low-rank regularization-based method for photon-limited image restoration

A photon-limited image can be represented as a pixel matrix limited by the relatively small number of collected photons. The image can also be seen as being contaminated by Poisson noise because the total number of photons follows the Poisson distribution. Through exploitation of the inherent properties of observation combined with application of a denoising method, an image can be significantly restored. In this paper, a hybrid clustering and low-rank regularization-based model (HCLR) is proposed based on the essential features of patch clustering and noise. An efficient Newton-type method is designed to optimize this biconvex problem. Experimental results demonstrate that HCLR achieves competitive denoising performance, especially for high noise levels, compared with state-of-the-art Poisson denoising algorithms.

A salt and pepper noise image denoising method based on the generative classification

In this paper, an image denoising algorithm is proposed for salt and pepper noise. First, a generative model is built on a patch as a basic unit and then the algorithm locates the image noise within that patch in order to better describe the patch and obtain better subsequent clustering. Second, the algorithm classifies patches using a generative clustering method, thus providing additional similarity information for noise repair and suppressing the interference of noise, abandoning those categories that consist of a smaller number of patches. Finally, the algorithm builds a non-local switching filter to remove the salt and pepper noise. Simulation results show that the proposed algorithm effectively denoises salt and pepper noise of various densities. It obtains a better visual quality and higher peak signal-to-noise ratio score than several state-of-the-art algorithms. In short, our algorithm uses a noisy patch as the basic unit, a patch clustering method to optimize the repair data set as well as obtain a better denoising effect, and provides a guideline for future denoising and repair methods.

Random Binary Local Patch Clustering Transforms Based Image Matching for Nonlinear Intensity Changes

This paper presents a new feature descriptor that is suitable for image matching under nonlinear intensity changes. The proposed approach consists of the following three steps. First, a binary local patch clustering transform response is employed as the transform space. The value of the new space exhibits a high similarity after changes in intensity. Then, a random binary pattern coding method extracts raw feature histograms from the new space. Finally, the discrimination of the proposed feature descriptor is enhanced by using a multiple spatial support region-based binning method. Experimental results show that the proposed method is able to provide a more robust image matching performance under nonlinear intensity changes.

Fission yeast Adf1 is necessary for reassembly of actin filaments into the contractile ring during cytokinesis

ADF/cofilin family proteins quickly disassemble actin in vitro, and are thought to be involved in various actin dynamics in the cell. Adf1 is a member of this family proteins expressed in fission yeast, and is thought to play roles in actin patch dynamics and also contractile ring formation during cytokinesis. We aimed to understand the function of this protein in cytokinesis in detail using the temperature-sensitive mutant adf1-1.Adf1 inactivation at a restrictive temperature during late G2 phase led to a clustering of actin patches at the cell ends. It was apparent that the inactivation occurred only in a few minutes. Furthermore, we found that the actin clusters migrated to the division site during anaphase possibly by the function of both myosin 5-1 and a myosin II. The migrated actin clusters, however, were not organized into the contractile ring. When Adf1 was inactivated at mid-anaphase B before contractile ring assembly, the ring was not formed, but it was formed when Adf1 was inactivated after this point. We conclude that Adf1 functions in the interphase actin dynamics and formation of the contractile ring during mitosis.

Visible and infrared image fusion using DTCWT and adaptive combined clustered dictionary

Getting the daylight information and the hidden target information in a single image is an active research topic in the domain of computer vision and image processing. In this paper, an image fusion technique, named as DTCWT-ACCD is proposed for the fusion of visible and infrared images. Firstly, an adaptive dictionary is constructed by combining several sub-dictionaries, learned from the clustered patches of source images. Then, the source images are decomposed by DTCWT to obtain the low frequency sub bands and high frequency sub bands. The low frequency sub bands are merged using a novel sparse based fusion rule while high frequency sub bands are combined using the maximum absolute value of coefficients with consistency verification (CV) check. Finally, the fused image is reconstructed by applying inverse DTCWT. The DTCWT-ACCD approach is experimentally tested with both subjective and objective evaluations to verify its competency. The results indicate that the DTCWT-ACCD approach is superior to conventional MST based methods and state-of-the-art sparse representation (SR) based methods.

A stochastic image denoising method based on adaptive patch-size

A new stochastic nonlocal denoising method based on adaptive patch-size is presented. The quality of restored image is improved by choosing the optimal nonlocal similar patch-size for each site of image individually. The method contains two phase. The first phase is to search the similar patches base on adaptive patch-size. The second phase is to design the denoising algorithm by making use of similar image patches obtained in the first step. The multiple clusters of similar patches for each pixel point are searched by using Markov-chain Monte Carlo sampling many times. Following, we adjust the patch-size according to the consistency of multiple clusters. This processing is repeated until we obtain the optimal patch-size and corresponding optimal patch cluster. We get the estimation of noise-free patch cluster by employing modified two-directional non-local method. Furthermore, the denoised image is obtained by using the method of superposition approach. The theoretical analysis and simulation results show that the method is feasible and effective.

Quasi-Zero-Voltage Controlled Etching of Macropores in n-Type Silicon

Here, we report for the first time on the controlled and accurate photoelectrochemical etching of macropores in n-type silicon in hydrofluoric acid-based electrolytes at anodic voltages below the electropolishing peak value down to quasi-zero Volt. Our results break a new ground on the controlled electrochemical dissolution of n-type silicon by further improving flexibility of this technology and pushing it towards commercial microelectronic applications. In the last few years the electrochemical micromachining (ECM) technology, which is based on the photoelectrochemical etching (PECE) of n-type silicon, is increasingly emerging as a promising silicon microstructuring technique alternative to standard both dry and wet etching tools1. A number of high aspect-ratio three-dimensional microstructures and microsystems have been reported over the past 10 years with applications ranging from automotive and biomedics to microelectronics and nanomedicine. The golden rule for the controlled PECE of macropores in low doped n-type silicon states to conduct the etching at anodic voltage values (Vetch ) above the electropolishing voltage value VPS (typically of a few Volts)2-5. Nonethless, even at Vetch >VPS there are a few constraints that limits the feasibility of macropores with arbitrary size and pitch. Specifically, it has been suggested that the maximum diameter of stable pores (about 10 μm for silicon of a few Ωcm) depends on the resistivity of the silicon substrate3,5. Moreover, the maximum pore pitch (about 20 μm for silicon of a few Ωcm) appears to be roughly limited to twice the depletion region at the silicon/electrolyte interface3. The starting material is a (100) oriented, n-type silicon wafer, with a resistivity of 3÷8 Ω∙cm and thickness of 750 μm, provided with a thermally grown 200-nm-thick silicon dioxide layer. A square lattice of holes with side of 4 μm and pitch of 20 μm was defined on the front surface of the silicon wafer by standard photolithography followed by buffered hydrofluoric acid (BHF) etching of the silicon dioxide and potassium hydroxide etching of silicon through a photoresist mask. Patterned wafers were electrochemically etched for 6000 s under back-side illumination (250 W halogen lamp) by using an aqueous solution of HF (concentration of 5% by vol). The current density/voltage (J/V) curve of the n-type silicon-electrolyte system under investigation has an electropolishing peak (VPS , JPS) at JPS =70.0 mA/cm2 (± 0.5 mA/cm2) and VPS =1.2 V. The same initial current density value Jetch0 =4.8 mA/cm2 was used for all the experiments. The Jetch value was then varied over time to ensure the etching of macropores with constant diameter over depth1,2,4. The anodic voltage values Vetch was varied among the different experiments from 0 to 2 V. Eventually, the electrochemically etched wafers were rinsed in ethanol and dried at 60 °C for 10 min. For each Jetch-Vetch parameter pair were performed 3 experiments, at least. Notice that, three of the Vetch values (i.e. 0, 0.4 and 0.8 V) are below the electropolishing peak value VPS and one is above (i.e. 2 V). Macropores etched at Vetch >VPS (i.e. 2 V) show a significant lateral branching and, in turn, a major roughness on the inner surface. This was somehow expected as both pitch (about 20 μm) and size (about 10 μm) are at close to the maximum features controlled by PECE according to the state-of-the-art literature3. On the contrary, the etching at Vetch <VPS led to very exciting and totally unexpected results, with the formation of regular macropores array without missing pores at any of tested voltages. Moreover, pore branching was significantly reduced as the anodic voltage was diminished below VPS and it completely disappeared at 0 V. A quantitative analysis of average values of etched macropores in terms of length and size points out that as the anodic voltage is reduced below the electropolishing value VPS regular macropores with reduced roughness, increased size, and shorter length are etched. We argue that the lower anodic voltage enables clustering of oxide patches at the tip of macropores over larger areas, compared to higher anodic voltages, thus slowing down the etching process in the vertical direction and increasing, in turn, the macropore diameter. Bassu, S. Surdo, L. M. Strambini, G. Barillaro, Advanced Functional Materials, 2012, 22 (6), 1222-1228.Föll, M. Christophersen, J. Carstensen, G. Hasse, Materials Science and Engineering R, 2002, 39(4), 93-142.Lehmann, Electrochemistry of Silicon: Instrumentation, Science, Materials, and Applications, Wiley-VCH, ISBN: 3-527-29321-3, Weinheim, 2002.Barillaro, F. Pieri, Journal of Applied Physics, 2005, 97, 116105-3.Barillaro, L.M. Strambini, Electrochemistry Communications, 2010, 12, 1314-1317.

A Homotopy Iterative Hard Thresholding Algorithm With Extreme Learning Machine for Scene Recognition

Vision-based scene recognition aims to find a semantic explanation of a scene image. First, how to learn a sparse feature representation by mapping each cluster of patches into few dimensions is one challenging issue. Sparse coding techniques with $l_{1}$ norm regularization are widely used to learn sparse features. However, $l_{1}$ norm regularization may not achieve enough sparsity. Furthermore, the initial value has a great influence on the sparsity of codes during the iterative optimization process. Second, how to learn a classifier in a supervised manner in order to improve the generalization is another issue. This paper therefore proposes a scene recognition method and it mainly includes two processes. One is the homotopy iterative hard thresholding (HIHT) algorithm that encodes the sparse representations of local patches by incorporating an $l_{0}$ norm regularization. Furthermore, a homotopy continuity strategy is used to improve the sparsity of feature codes by adaptively tuning the regularization factor from large to small values and using the sparse solution of last iteration as the warm start of the next iteration. The other is the extreme learning machine (ELM)-based classifier. Experimental results in 15-class scene data set have shown that the HIHT algorithm outperforms other unsupervised sparse feature learning algorithms in terms of sparsity and entropy. Meanwhile, the ELM-based scene recognition method outperforms the other state-of-the art methods in terms of recognition accuracy.

Topology optimization for concurrent design of structures with multi-patch microstructures by level sets

This paper focuses on the novel concurrent design for cellular structures consisting of multiple patches of material microstructures using a level set-based topological shape optimization method. The macro structure is featured with the configuration of a cluster of non-uniformly distributed patches, while each patch hosts a number of identical material microstructures. At macro scale, a discrete element density based approach is presented to generate an overall structural layout involving different groups of discrete element densities. At micro scale, each macro element is regarded as an individual microstructure with a discrete intermediate density. Hence, all the macro elements with the same discrete densities (volume fractions) are represented by a unique microstructure. The representative microstructures corresponding to different density groups are topologically optimized by incorporating the numerical homogenization approach into a parametric level set method. The multiscale concurrent designs are integrated into a uniform optimization procedure, so as to optimize both topologies for the macrostructure and its microstructures, as well as locations of the microstructures in the design space. Numerical examples demonstrate that the proposed method can substantially improve the structural performance with an affordable computation and manufacturing cost.

Deep linear autoencoder and patch clustering-based unified one-dimensional coding of image and video

This paper proposes a unified one-dimensional (1-D) coding framework of image and video, which depends on deep learning neural network and image patch clustering. First, an improved K-means clustering algorithm for image patches is employed to obtain the compact inputs of deep artificial neural network. Second, for the purpose of best reconstructing original image patches, deep linear autoencoder (DLA), a linear version of the classical deep nonlinear autoencoder, is introduced to achieve the 1-D representation of image blocks. Under the circumstances of 1-D representation, DLA is capable of attaining zero reconstruction error, which is impossible for the classical nonlinear dimensionality reduction methods. Third, a unified 1-D coding infrastructure for image, intraframe, interframe, multiview video, three-dimensional (3-D) video, and multiview 3-D video is built by incorporating different categories of videos into the inputs of patch clustering algorithm. Finally, it is shown in the results of simulation experiments that the proposed methods can simultaneously gain higher compression ratio and peak signal-to-noise ratio than those of the state-of-the-art methods in the situation of low bitrate transmission.

Optical coherence tomography image denoising using Gaussianization transform.

We demonstrate the power of the Gaussianization transform (GT) for modeling image content by applying GT for optical coherence tomography (OCT) denoising. The proposed method is a developed version of the spatially constrained Gaussian mixture model (SC-GMM) method, which assumes that each cluster of similar patches in an image has a Gaussian distribution. SC-GMM tries to find some clusters of similar patches in the image using a spatially constrained patch clustering and then denoise each cluster by the Wiener filter. Although in this method GMM distribution is assumed for the noisy image, holding this assumption on a dataset is not investigated. We illustrate that making a Gaussian assumption on a noisy dataset has a significant effect on denoising results. For this purpose, a suitable distribution for OCT images is first obtained and then GT is employed to map this original distribution of OCT images to a GMM distribution. Then, this Gaussianized image is used as the input of the SC-GMM algorithm. This method, which is a combination of GT and SC-GMM, remarkably improves the results of OCT denoising compared with earlier version of SC-GMM and even produces better visual and numerical results than the state-of-the art works in this field. Indeed, the main advantage of the proposed OCT despeckling method is texture preservation, which is important for main image processing tasks like OCT inter- and intraretinal layer analysis. Thus, to prove the efficacy of the proposed method for this analysis, an improvement in the segmentation of intraretinal layers using the proposed method as a preprocessing step is investigated. Furthermore, the proposed method can achieve the best expert ranking between other contending methods, and the results show the helpfulness and usefulness of the proposed method in clinical applications.

A feature descriptor based on the local patch clustering distribution for illumination-robust image matching

This paper proposes a feature descriptor based on the local patch clustering distribution (LPCD), which preserves the salient features of a given image following changes in illumination. To mitigate the effects of illumination change, the proposed LPCD methodology consists of two steps. First, a local patch clustering assignment map is constructed by pairing the source image with a reference image. To resolve the quantization problem caused by an illumination change, a dual-codebook clustering method is employed so that an effective local patch clustering feature space can be constructed. Second, in the feature encoding process, the impact of the informative local patches that contain textural information is enhanced when using a saliency detection response as a method of weighting every local patch when the histogram feature is extracted. Experimental results show that the proposed local patch clustering space is more robust than the conventional intensity order-based space in response to changes in illumination.

Gaussian mixture image restoration based on maximum correntropy criterion

Using Gaussian mixture model (GMM) as a patch prior has achieved great success in image restoration (IR). However, based on the derivations, the objective function of the conventional GMM-based IR methods can be expressed as the minimum of mean square error (MMSE) criterion based optimisation problem. As outliers may exist in the patch cluster obtained even within a spatially constrained window and MMSE is sensitive to outliers, using the MMSE criterion based strategy would lead to error results. A maximum correntropy criterion (MCC) based IR method with GMM as a patch prior is proposed. The proposed method can automatically identify outliers and assign weight for each patch within the k-nearest-neighbour patches with respect to each exemplar patch in the image and thus can robustly estimate the Gaussian parameters which results in more accurate estimation of the image patches. Finally, an effective iterative optimisation algorithm is designed to solve the proposed objective function under the MCC criterion. The experimental results for image inpainting demonstrate the capabilities of the proposed method.

Image restoration using structured sparse representation with a novel parametric data-adaptive transformation matrix

In this paper, we propose a new framework for image restoration by combining nonlocal regularization technique with structured sparse representation using a novel parametric data-adaptive transformation matrix. Specifically, we first present a parametric nonlocal difference operator in the context of nonlocal regularization with the insight that the weight functions in the definition of the nonlocal difference operator play an important role in modeling image prior information. The proposed difference operator not only gives flexibility in modeling prior knowledge of clean images compared with the conventional methods, but also leads to a novel parametric data-adaptive transformation matrix. Then, we propose a structured sparse representation model based on this new transformation matrix for image restoration, taking advantage of the sparse nature of the transform coefficients of image patches over the corresponding transformation matrices. Unlike most structured sparse models, our suggested method does not require the nonlocal self-similarity assumption or patch clustering algorithms to find similar grouped patches either in the image under test or in an external dataset. Finally, an effective optimization algorithm is designed to solve the corresponding sparse inverse problem. Extensive experimental comparisons with state-of-the-art image deblurring and super-resolution algorithms validate the effectiveness of our proposed method.