Abstract
Due to the characteristics of hyperspectral images (HSIs), such as their high spectral resolution and multiple continuous narrow bands, HSI technology has become widely used in fields such as target recognition, environmental detection, and agroforestry detection. HSIs are subject, for various reasons, to noise in the processes of data acquisition and transmission. Therefore, the denoising of HSIs is very necessary and important. In this article, according to the characteristics of HSIs, an HSI denoising model combining principal component analysis (PCA) and CANDECOMP/PARAFAC decomposition (CP decomposition) is proposed, which is called PCATensorDecomp. First, we use PCA to reduce the dimension of HSI signals by obtaining the first K principal components and get the principal composite components. The lowrank part corresponding to the first K principal components is considered the characteristic signal. Then, lowrank CP decomposition is carried out, to denoise the first principal components and the remaining minor components, the secondary composite components, which contain a large amount of noise. Finally, the inverse PCA is then used to restore the HSIs denoised, such that the effect of comprehensive denoising is achieved. To test the effectiveness of the improved algorithm introduced in this article, we compare it with several methods on simulated and real hyperspectral data. The results of the analysis herein indicate that the proposed algorithm possesses a good denoising effect.
Graphical abstract
1 Introduction
In recent years, the rapid development of hyperspectral image (HSI) technology has provided abundant spectral spatial features for a wide range of applications, such as target detection and classification [1]. An HSI is equivalent to the superposition of many grayscale images. Therefore, the HSI can be regarded as a cube structure image, which contains rich information. On one hand, it is equivalent to a twodimensional spatial image, from which spatial dimension information can be obtained. On the other hand, it also contains the spectral structure of the observed target.
In the process of imaging and transmission, HSIs are inevitably polluted by noise from sensors, the atmosphere, and other sources, such that the quality of HSIs may be seriously reduced, which undoubtedly has adverse effects on the extraction of HSI information. Therefore, as a key preprocessing step to improve the subsequent recognition ability, HSI noise reduction has attracted the attention of many scholars and has become a hot topic in the field of hyperspectral remote sensing in recent years. Compared with panchromatic and multispectral imagery, the noise distribution of hyperspectral imagery is more complex. As a result, traditional gray image denoising methods, such as KSVD [2], block matching, and 3D filtering [3], may not always be good at removing noise from HSIs. Therefore, a reasonable and effective denoising method should be designed, according to the characteristics of HSIs.
In recent years, numerous denoising methods for HSIs have been proposed, most of which have focused on enhancing the restoration effect in the spatial and spectral domains. Denoising methods can be roughly divided into two types: The first type refers to twodimensional image denoising methods, which introduces the spatial spectrum characteristics in different transform domains for HSI denoising, such as wavelet transforms [4], principal component analysis (PCA) [5], and sparse 3D transformdomain collaborative filtering (BM3D) method [3]. However, such methods only consider spatial domain denoising, and they do not use spectral information to attenuate noise. Thus, there generally still exists some noise, which could not be removed by such methods. HSI denoising must consider both spatial and spectral domains. The second type of methods consists of denoising in the spectral and spatial domains in different stages. To better reduce the noise in homogeneous regions while maintaining edge and texture information, Yuan et al. [6] have proposed a spectral adaptive total variation model to smooth the noise and signals. In addition, Sun and Luo [7] have proposed a threedimensional mixed denoising method in the differential domain of HSIs, in which a spectral dimension transformation is first performed on the HSI, following which the inverse transformation is performed after denoising. Compared with the first type, methods of the second type take both spatial and spectral information into consideration in the denoising algorithm, such that the denoising effect can be significantly improved.
An HSI contains three dimensions: one spectral dimension and two spatial dimensions. It can be modeled as threeorder tensor data. Therefore, tensorbased models can improve the denoising results, because they can fully capture the spatial–spectral correlation of the HSI. Recently, with the development of tensor theory, many tensorbased HSI denoising methods have emerged, such as the Tucker decompositionbased HSI denoising algorithm [8,9], but these methods often have the problem of low efficiency and ignore the nonlocal selfsimilarity prior. Zheng et al. [10] proposed a novel nonlocal patchbased FCTN (NLFCTN) decomposition for MSI inpainting. At the same time, due to the characteristics of HSIs, the structured highdimensional information inherent in the original observations may be discarded when the highdimensional HSI data are converted into twodimensional data by traditional methods and processed separately. Inspired by the noise estimation method proposed in ref. [11], Meng et al. [12] proposed an HSI denoising method by jointly exploiting Tucker tensor decomposition and PCA based on the noise power ratio (NPR) analysis. With the NPR, they estimate the Tucker rank. And as this method exploits the correlation effects in all of the dimensions of HSI data, it gets some good results when denoising.
When using Tucker decomposition, the first step is to estimate the rank of Tucker decomposition. However, for a fixed Nrank image, the uniqueness of the Tucker decomposition is not guaranteed. Scholars have made some improvements on this basis, such as the kernel nonnegative Tucker decomposition [13], nonnegative Tucker decomposition based on nonlocal selfsimilarity in the spatial domain [14], and Tucker decomposition based on nonlocal selfsimilarity and global correlation across the spectrum [15]. Although these methods can better demonstrate the structural correlation of HSIs and reduce the artifacts of overlapping regions, they cannot solve the problem that the Tucker decomposition is not unique.
CANDECOMP/PARAFAC decomposition (CP decomposition), can be regarded as a special form of Tucker decomposition [16,17]. It implements the tensor approximation by manipulating the CP rank; thus, it has been used as an effective method. Liu et al. [18] proposed an HSI denoising method based on CANDECOMP/PARAFAC decomposition (CP decomposition) and estimated the CP rank depending on the variance of additive noise. However, in practice, the variance is often unknown. Therefore, some loworder models of joint local and nonlocal selfsimilarity can be used to improve the quality of denoising, such as the nonlocal image denoising model based on the weighted tensor rank1 decomposition of Wu et al. and Guo et al. [19,20].
In addition to being widely used in the field of HSI denoising [21,22], lowrank priors can improve HSI superresolution. Dian and Li [23] proposed a novel subspacebased low tensor multirank regularization method for the fusion, which makes full use of spectral correlations and nonlocal similarities of the HRHSI. Xue et al. [24] propose a new HSI superresolution method that fully considers the spatial/spectral subspace lowrank relationships between available highresolution multispectral image/lowresolution HSI and potential HSI.
Another widely used HSI denoising method is tensor regularization [25,26]. Xue et al. [25] used the local regularization CP lowrank tensor decomposition method for HSI denoising. Zhang et al. [27] approximated the rank by a sum of tensor ranks in different directions, considering the validity of singular value decomposition (SVD) in the literature. Liu et al. [28] suggest a novel HSI restoration model by introducing a fibered rank constrained tensor restoration framework with an embedded plugandplaybased regularization. For HIS mixed noise, Zheng et al. [29] proposed a doublefactorregularized LRTF model for HSI mixed noise removal. Zheng et al. [30] proposed a fibered rank minimization model for HSI mixed noise removal, in which the underlying HSI is modeled as a lowfiberedrank component. The key problem in tensor regularization is how to construct a reasonable regularization term to reflect the global correlation and nonlocal similarity. Tensor regularization can introduce prior information into the HSI denoising and retain the spectral and spatial dimensional information. It can be seen that tensor regularization has a good application prospect for HSI denoising.
To make use of the advantage of tensor decomposition and tensor regularization and take the characteristics of HSIs into account, in this article, we propose an HSI denoising model combining PCA and CP decomposition, which we call PCA TensorDecomp.
In summary, the main contributions of this work are as follows:
A novel method proposed to solve HSI denoising by combining PCA and CP decomposition named PCA TensorDecomp.
Inspired by the alternating direction method of multipliers (ADMM) and alternate least squares, a new algorithm has been proposed to effectively solve the method proposed in this article.
A broad experimental analysis has been provided based on several datasets. The numerical outcomes are also corroborated by a qualitative analysis.
The remainder of this article is arranged as follows: Section 2 briefly introduces the principles of PCA and CP decomposition. In Section 3, the PCA TensorDecomp is proposed. In Section 4, some numerical experiments are provided and Section 5 is the conclusion.
2 Notations and related methods
2.1 Notations
In this article, lowercase or uppercase letters (e.g.,
2.2 PCA
HSI data are complex. HSIs consist of dozens or even hundreds of continuous and subdivided spectral bands imaging the target region at the same time, which provides us with abundant information and is conducive to the fine division of ground objects. The increase in band number not only leads to information redundancy but can also lead to a requirement for rather complex data processing, such that it is necessary to reduce the dimension of the HSI data. PCA can separate the principal and secondary components of HSIs well. On one hand, redundant data can be removed; on the other hand, essential information can be retained. It is, therefore, suitable for the complex data processing of HSIs.
The essence of PCA is to transform a group of related variables into a group of unrelated variables through an orthogonal transformation and it has been widely applied in many fields [5,31]. The formula is as follows:
where X is the original data,
Considering that PCA cannot directly process threedimensional data, such as HSIs, we carry out a preprocessing to reduce the dimension before conducting the PCA. The specific steps for dimensionality reduction of hyperspectral data are as follows.
The first step is to convert the HSI data
The second step is to calculate the covariance matrix S of the data matrix X. The specific formula is as follows:
where
In the third step, with the covariance matrix S obtained, and in the second step, solve the eigenequation
Finally, all the eigenvectors are sequentially arranged to form the orthogonal matrix A.
Usually, with PCA of HSIs, the new principal composite components contain less and less information. Therefore, we mainly retain most of the relevant information with the first few principal components and discard the secondary composite components, which are the socalled noise signal components. However, those secondary composite components usually contain some useful information and, so, after discarding the secondary composite components, the reconstruction of HSIs will lead to some loss of the information. Therefore, how to extract information from the secondary composite components is a very critical problem.
2.3 CP decomposition
The essence of CP decomposition is to decompose thirdorder tensor data into a sum of R rank1 tensors [16]. Note that HSIs can be treated as cube data (namely, a thirdorder tensor). Because of this property, CP decomposition can not only solve the problem of lowrank tensor completion [32] but also be used for HIS denoising.
Given a thirdorder tensor
where R is an integer and
Tensor
where
In general, the column vectors of the matrices A, B, and C are normalized. Let
where
CP decomposition requires two steps:
Determine the number of decomposition ranks; that is, the rank of the socalled tensor.
The rank of a tensor is different from that of a matrix, such that there is no simple method to solve the rank of CP decomposition. That is to say, solving the tensor rank is currently an NPhard problem, and there is no definite method or algorithm, which is accurate for determining the decomposition rank. We can, however, choose several different values to compare their effect, or start from rank1 and carry out an iterative approach. Of course, the tensor ranks of some dimensions have been obtained by some scholars, which provide us with a good reference in particular cases [33,34].
Once R, the rank of a tensor, has been determined, decomposition of the tensor can be performed.
When we solve the CP decomposition of a thirdorder tensor, it is equivalent to using the sum of R rank1 tensors to approximate the original tensor
where
Equation (7) can be solved by the ADMM [35]. The optimization principle of the ADMM optimization algorithm is to first fix the other parameters, except for the one to be optimized. In the above formula, the specific process is fixed first and then optimized. The operation is repeated until the iteration has reached a certain number or the error gets to the requirement of the experiment. The specific optimization processes of different parameters are similar, so the following only introduces the optimization process for the parameter A.
As all of the matrices are fixed, except for one, the problem becomes a linear leastsquares problem. Here, we optimize for parameter A, with B and C fixed. According to equation (5), we can write the above minimization problem, in the matrix form, as:
When
Here
Due to the special form of the Khatri–Rao product, the above formula can be written as:
The advantage of equation (10) is that we only need to calculate the pseudoinverse of a
Algorithm 1: ADMMbased CP denoising algorithm
Input: noisy image
Output: denoising image denoising image
1. Initialization:
2. while
3. Update
4. Update
5. Update
6. Update
7. end
For the parameters
The corresponding denoising algorithm is summarized in Algorithm 1, in which,
3 PCA tensor decomposition method
HSI data contain multidimensional information, such that analysis of the data in each dimension may lose the potential structure of the data. Fortunately, tensor decomposition can effectively mine the potential information among the data. In this section, we propose an HSI denoising method based on the PCA and lowrank CP decomposition (PCA TensorDecomp).
The denoising process of PCA TensorDecomp is shown in Figure 1.
The HSI model is formulated as:
where
Considering the roles of PCA, on one hand, it can achieve the aim of dimensionality reduction for the HSI and, on the other hand, it can separate the principal composite components from the secondary composite components. With the few main principal composite components, most of the relevant information can be retained. As mentioned earlier, those secondary composite components contain most of the noise, but they still have some useful information. Discarding the secondary composite components, this useful information will lose.
As the HSIs are tensortype data, lowrank CP decomposition can be used for denoising. The use of a lowrank tensor decomposition can effectively remove the noise contained in the tensor.
Thus, to combine the advantages of PCA and CP decomposition, we propose a denoising algorithm based on the PCA and lowrank tensor CP decomposition named PCA TensorDecomp.
The steps of the algorithm are as follows:
Step 1: Input the noisy image
Step 2: Preprocess the hyperspectral noise image
Step 3: Separate the principal and minor components and keep the previous principal composite components unchanged, denoted as
Step 4: Denoise the previous principal composite components
where the symbol definition above is the same as in Section 2.2.
Step 5: Finally, combine the denoising former principal composite component
where
Step 6: Finally, the inversely converted data
The flow chart of the proposed algorithm is shown below.
Algorithm 2: PCA TensorDecomp denoising algorithm
Input: noisy image
Output: denoising image denoising image
1. Preprocess
2. Separate the principal components
3.
4. Combine
5. Inversely transform
6. Restore
4 Experimental results and analysis
To test the effectiveness of the proposed denoising algorithm, in this section, we mainly report the results of experiments on simulated and real HSI data. Five traditional HSI denoising algorithms, Tucker decomposition [38], Adaptive NonLocal Means 3D (ANLM3D) [39], lower rank tensor approximation (LRTA) [40], truncated singular value decomposition (tSVD) [41], and PARAFAC [18], were selected for comparison. We evaluated the denoising results both subjectively and objectively. That is, we discuss the merits and demerits of five denoising algorithms from the perspective of both visual qualitative analysis and quantitative analysis. Peak signaltonoise ratio (PSNR), structural similarity (SSIM), feature similarity (FSIM) [42], and erreur relative globale adimensionnelle de synthese (ERGAS) [43] were used for the quantitative analysis of simulated HSIs and real hyperspectral data.^{[1]} The values of SSIM and FSIM are in the range [0,1] while the values of PSNR approach infinity as a limit. The greater the value, the less image distortion. Different from the former three measures, the smaller ERGAS is, the better does the target HSI estimate the reference one. Meanwhile, we then examine the horizontal average fluctuation of a given band. Notice that the parameter setting for PCA was three in all the following experiments.
4.1 Simulated HSI denoising experiments
We selected the Washington DC Mall set from the Hyperspectral Digital Imagery Collection Experiment sensor for the simulation experiments. The original image is of size
During the experiment, the number of ranks for the first principal components for CP decomposition was 240/22 (denoised with noise intensity 0.1/denoised with noise intensity 0.2) and the number of ranks for the second principal components for CP decomposition was 180/4.
Comparing Figures 2 and 3, it can be seen that the tSVD algorithm had the worst effect in both cases because it used matrix factorization technology to perform the denoising processing. The image restored by ANLM3D is blurred and has artifacts. Although Tucker obtained acceptable results at low noise levels, it could not remove noise at high noise levels. LRTA and PARAFAC effectively eliminated noise, but image detail was lost. These visual conditions can be observed in the enlarged box. In comparison, the PCA TensorDecomp algorithms achieved superior results competitive with the other methods.
Table 1 presents the quantitative results from comparison of the denoising approaches with the Washington DC Mall set at two different noise levels. We highlight the top two algorithms with two different representations of the data. Between the two, the best performance for each quality index is highlighted by bolding the data, and the secondbest performance is highlighted by highlighting the data. It can also be seen, from the comparison results in Table 1 that the average PSNR of the denoising method proposed in this article was the highest, indicating that it achieves a better improvement of the denoising results, compared with the other denoising algorithms. At the same time, the structural similarity of the denoising algorithm mentioned in this article was also significantly improved, compared with that of the other five algorithms, which indicates that the denoising algorithm proposed in this article also significantly improved the authenticity of image restoration.
Noise intensity  Method  MPSNR  MSSIM  MFSIM  MERGAS 

0.1  Noisy image  20.00  0.432  0.732  422.499 
Tucker  32.83  0.920  0.953  93.569  
ANLM3D  27.95  0.740  0.838  162.521  
LRTA  33.62  0.929  0.959  85.635  
tSVD  26.03  0.690  0.867  209.735  
PARAFAC  31.59  0.902  0.942  108.098  
PCA TensorDecomp  35.82  0.953  0.975  67.887  
0.2  Noisy image  13.98  0.191  0.560  844.942 
Tucker  29.47  0.877  0.929  139.550  
ANLM3D  25.78  0.630  0.781  208.236  
LRTA  30.05  0.858  0.919  128.480  
tSVD  21.36  0.461  0.764  359.067  
PARAFAC  30.43  0.875  0.929  123.041  
PCA TensorDecomp  32.14  0.879  0.940  115.865 
Figure 4 shows the PSNR and SSIM values of each band on the Washington DC Mall set at two different noise levels. From the experimental results in Figure 4, we can see that, with different denoising algorithms, the PSNRs and SSIMs of different subimages from the simulation experiment were all greater, compared to the noisy image. On the other hand, for different noise intensities and subimages, the PCA TensorDecomp denoising algorithm was more effective than most of the other denoising methods because it obtained the best results in most bands. This indicates that the PCA TensorDecomp algorithm proposed in this article can outperform the traditional denoising algorithm.
4.2 Real HSI denoising experiments
In this experiment, two groups of real HSI data were selected, namely Indian Pines and Gaofen5 data.
4.2.1 Indian Pines data
The real data were collected using the AVIRIS sensor at the Indian Pines Test site in northwestern Indiana. The size of the HSI data is
The HSI was denoised and the three denoising algorithms (Tucker Decomposition, PARAFAC, ANLM3D) mentioned in the previous section were compared. In the experiment, the rank of CP decomposition was 20/20. In the consideration of analyzing the experimental results qualitatively, row signature curves of the first band estimated by all the compared methods were compared in Figure 6.
Comparing Figure 5, it can be seen that the denoising effect of PARAFAC was not very good for real data denoising. Tucker can get clear edges that ANLM3D cannot get, but it has artifacts in some areas.
Figure 6 shows the horizontal mean profiles of band 1 before and after denoising. As shown in Figure 6a, due to the existence of noise, there are rapid fluctuations in the curve before denoising. After processing by several denoising algorithms, the fluctuations are more or less suppressed. Here, we can see that the curve of the proposed PCA TensorDecomp method is more stable, which is in accordance with the visual results presented in Figure 5.
4.2.2 Gaofen5 data
To test the effectiveness of the denoising algorithm proposed in this article, we adopted another data set Gaofen5 data. The original data set consists of
As above, different denoising algorithms were adopted to denoise the Gaofen5 data. Results are shown in Figure 7. As you can see from the figure, Tucker gets a distorted image and ANLM3D gets a blurred image. Figure 8 shows the horizontal mean profiles of the third band before and after denoising. As shown in Figure 8a, we get the same conclusion as the Indian Pines data set. Here, we can see that compared with other denoising algorithms, PCA TensorDecomp proposed in this article shows more superior.
4.3 Analysis of CP rank
In the PCA TensorDecomp model, when executing the algorithm, we need to determine the size of the CP rank to get the best denoising effect. Figure 9 shows the changes in PSNR corresponding to different CP ranks in WDC datasets under two noise intensities. It can be seen from the figures that when the noise intensity is 0.1, the greater the CP rank of the first principal component and the CP rank of the second principal component is in the interval of [22,30], the better the denoising effect of the proposed algorithm. When the noise intensity is 0.2, the CP rank of the first principal component is in the interval of [170, 190], and the smaller the CP rank of the second principal component, the better the denoising effect of the proposed algorithm. Therefore, the parameter size set in this article is reasonable and benign.
5 Conclusion
A new method for HSI noise reduction using PCA and a lowrank CP decomposition model (PCA TensorDecomp) was presented in this article. This method takes advantage of the fact that PCA can reduce the complexity of the processing data, as well as take into account the characteristics of HSIs and the correlations between various bands. To verify the validity of this algorithm, we selected simulated HSI data and two sets of real SIs to carry out comparative experiments, through both qualitative (i.e., assessing the visual effect) and quantitative analyses of the proposed method and five traditional methods. The results demonstrated that, compared with these methods, the proposed PCA TensorDecomp method can effectively reduce the noise of HSIs while retaining their fine structure. There are some aspects to be improved; for example, in the process of PCA, the principal component selection is carried out manually according to the degree of contribution. Meanwhile, a better algorithm should be sought, to get the number of CP decomposition ranks.
Acknowledgments
This work was supported in part by the Sichuan Science and Technology Program under Grant 2021YJ0351 and Grant 2021YFG0319, in part by the Opening Fund of Geomathematics Key Laboratory of Sichuan Province under Grant scsxdz2019zd03 and Grant scsxdz2021yb02, and in part by the National Key Research and Development Program of China under Grant 2017YFC0601505.

Author contributions: Hao Wu, Youhua Wei, Ruihan Yue, and Jun Feng proposed the model. Ruixue Gao and Rui Wen designed the experiments and carried them out. Ruixue Gao prepared the manuscript with contributions from all coauthors. All authors revised the manuscript.

Conflict of interest: Authors state no conflict of interest.
References
[1] BioucasDias JM, Plaza A, Dobigeon N, Parente M, Du Q, Gader P, et al. Hyperspectral unmixing overview: Geometrical, statistical, and sparse regressionbased approaches. IEEE J Sel Top Appl Earth Obs Remote Sens. 2012;5(2):354–79. 10.1109/JSTARS.2012.2194696.Search in Google Scholar
[2] Aharon M, Elad M, Bruckstein A. KSVD: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Trans Signal Process. 2006;54(11):4311–22. 10.1109/TSP.2006.881199.Search in Google Scholar
[3] Dabov K, Foi A, Katkovnik V, Egiazarian K. Image denoising by sparse 3D transformdomain collaborative filtering. IEEE Trans Image Process. 2007;16(8):2080–95. 10.1109/TIP.2007.901238.Search in Google Scholar
[4] Mallat S. A wavelet tour of signal processing. 3rd edn. Boston: Academic Press; 2009.Search in Google Scholar
[5] Huo LG, Feng XC. Denoising of hyperspectral remote sensing image based on principal component analysis and dictionary learning. Dianzi Yu Xinxi Xuebao/J Electron Inf Technol. 2014;36:2723–9. 10.3724/SP.J.1146.2013.01840.Search in Google Scholar
[6] Yuan Q, Zhang L, Shen H. Hyperspectral image denoising employing a spectralspatial adaptive total variation model. IEEE Trans Geosci Remote Sens. 2012;50(10 PART1):3660–77. 10.1109/TGRS.2012.2185054.Search in Google Scholar
[7] Sun L, Luo JS. Threedimensional hybrid denoising algorithm in derivative domain for hyperspectral remote sensing imagery. Guang Pu Xue Yu Guang Pu Fen Xi/Spectroscopy Spectr Anal. 2009;29(10):2717–20. 10.3964/j.issn.10000593(2009)10271704.Search in Google Scholar
[8] Xie Q, Zhao Q, Meng D, Xu Z. kroneckerbasisrepresentation based tensor sparsity and its applications to tensor recovery. IEEE Trans Pattern Anal Mach Intell. 2018;40(8):1888–902. 10.1109/TPAMI.2017.2734888.Search in Google Scholar PubMed
[9] Fan H, Li C, Guo Y, Kuang G, Ma J. Spatialspectral total variation regularized lowrank tensor decomposition for hyperspectral image denoising. IEEE Trans Geosci Remote Sens. 2018;56(10):6196–213. 10.1109/TGRS.2018.2833473.Search in Google Scholar
[10] Zheng WJ, Zhao XL, Zheng YB, Pang ZF. Nonlocal patchbased fullyconnected tensor network decomposition for multispectral image inpainting. IEEE Geosci Remote Sens Lett. 2021;19:1–5. 10.1109/LGRS.2021.3124804.Search in Google Scholar
[11] BioucasDias JM, Nascimento JMP. Hyperspectral subspace identification. IEEE Trans Geosci Remote Sens. 2008;46(8):2435–45. 10.1109/TGRS.2008.918089.Search in Google Scholar
[12] Meng S, Huang LT, Wang WQ. Tensor decomposition and PCA jointed algorithm for hyperspectral image denoising. IEEE Geosci Remote Sens Lett. 2016;13(7):897–901. 10.1109/LGRS.2016.2552403.Search in Google Scholar
[13] Karami A, Yazdi M, Asli AZ. Noise reduction of hyperspectral images using kernel nonnegative tucker decomposition. IEEE J Sel Top Signal Process. 2011;5(3):487–93. 10.1109/JSTSP.2011.2132692.Search in Google Scholar
[14] Bai X, Xu F, Zhou L, Xing Y, Bai L, Zhou J. Nonlocal similarity based nonnegative tucker decomposition for hyperspectral image denoising. IEEE J Sel Top Appl Earth Obs Remote Sens. 2018;11(3):701–12. 10.1109/JSTARS.2018.2791718.Search in Google Scholar
[15] Xie Q, Zhao Q, Meng DY, Xu ZB, Gu SH, Zuo WM, et al. Multispectral images denoising by intrinsic tensor sparsity regularization. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition; 2016. Vol. 2016Decem. p. 1692–700. 10.1109/CVPR.2016.187.Search in Google Scholar
[16] Carroll JD, Chang JJ. Analysis of individual differences in multidimensional scaling via an nway generalization of ‘EckartYoung’ decomposition. Psychometrika. 1970;35(3):283–319. 10.1007/BF02310791.Search in Google Scholar
[17] Harshman RA. Foundations of the PARAFAC procedure: Models and conditions for an ‘explanatory’ multimodal factor analysis. Vol. 16. University Microfilms, Ann Arbor, Michigan: UCLA Working Papers in Phonetics, No. 10,085; 1970. p. 1–84.Search in Google Scholar
[18] Liu X, Bourennane S, Fossati C. Denoising of hyperspectral images using the parafac model and statistical performance analysis. IEEE Trans Geosci Remote Sens. 2012;50(10):3717–24. 10.1109/TGRS.2012.2187063.Search in Google Scholar
[19] Wu Y, Fang L, Li S. Weighted tensor rank1 decomposition for nonlocal image denoising. IEEE Trans Image Process. 2019;28(6):2719–30. 10.1109/TIP.2018.2889914.Search in Google Scholar PubMed
[20] Guo X, Huang X, Zhang L, Zhang L. Hyperspectral image noise reduction based on rank1 tensor decomposition. ISPRS J Photogramm Remote Sens. 2013;83:50–63. 10.1016/j.isprsjprs.2013.06.001.Search in Google Scholar
[21] Zeng H, Xie X, Cui H, Yin H, Ning J. Hyperspectral image restoration via global L 1–2 spatial–spectral total v ariation regularized local lowrank tensor recovery. IEEE Trans Geosci Remote Sens. 2020;59(4):3309–25.10.1109/TGRS.2020.3007945Search in Google Scholar
[22] Zeng H, Xie X, Cui H, Zhao Y, Ning J. Hyperspectral image restoration via cnn denoiser prior regularized lowrank tensor recovery. Computer Vis Image Underst. 2020;197:103004.10.1016/j.cviu.2020.103004Search in Google Scholar
[23] Dian R, Li S. Hyperspectral image superresolution via subspacebased low tensor multirank regularization. IEEE Trans Image Process. 2019;28(10):5135–46.10.1109/TIP.2019.2916734Search in Google Scholar PubMed
[24] Xue J, Zhao YQ, Bu Y, Liao W, Chan JC, Philips W. Spatialspectral structured sparse lowrank representation for hyperspectral image superresolution. IEEE Trans Image Process. 2021;30:3084–97.10.1109/TIP.2021.3058590Search in Google Scholar PubMed
[25] Xue J, Zhao Y, Liao W, Chan JCW. Nonlocal lowrank regularized tensor decomposition for hyperspectral image denoising. IEEE Trans Geosci Remote Sens. 2019;57(7):5174–89. 10.1109/TGRS.2019.2897316.Search in Google Scholar
[26] Kervrann C, Boulanger J. Optimal spatial adaptation for patchbased image denoising. IEEE Trans Image Process. 2006;15(10):2866–78. 10.1109/TIP.2006.877529.Search in Google Scholar
[27] Zhang Z, Ely G, Aeron S, Hao N, Kilmer M. Novel methods for multilinear data completion and denoising based on tensorSVD. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition; 2014. p. 3842–9. 10.1109/CVPR.2014.485.Search in Google Scholar
[28] Liu YY, Zhao XL, Zheng YB, Ma TH, Hongyan Z. Hyperspectral image restoration by tensor fibered rank constrained optimization and plugandplay regularization. IEEE Trans Geosci Remote Sens. 2021;60:1–17. Accepted. 10.1109/TGRS.2020.3045169.Search in Google Scholar
[29] Zheng YB, Huang TZ, Zhao XL, Chen Y, He W. doublefactorregularized lowrank tensor factorization for mixed noise removal in hyperspectral image. IEEE Trans Geosci Remote Sens. 2020;58(12):8450–64.10.1109/TGRS.2020.2987954Search in Google Scholar
[30] Zheng YB, Huang TZ, Zhao XL, Jiang TX, Ma TH, Ji TY. Mixed noise removal in hyperspectral image via lowfiberedrank regularization. IEEE Trans Geosci Remote Sens. 2020;58(1):734–49.10.1109/TGRS.2019.2940534Search in Google Scholar
[31] Zhou B, Duan X, Ye D, Wei W, Woźniak M, Damaševičius R. Heterogeneous image matching via a novel feature describing model. Appl Sci. 2019;9(22):4792.10.3390/app9224792Search in Google Scholar
[32] Xue J, Zhao Y, Huang S, Liao W, Chan JCW, Kong SG. Multilayer sparsitybased tensor decomposition for lowrank tensor completion. IEEE Trans Neural Netw Learn Syst. 2021;1–15. 10.1109/TNNLS.2021.3083931.Search in Google Scholar PubMed
[33] Hitchcock FL. The expression of a tensor or a polyadic as a sum of products. J Math Phys. 1927;6(1–4):164–89. 10.1002/sapm192761164.Search in Google Scholar
[34] Kruskal JB. Threeway arrays: rank and uniqueness of trilinear decompositions, with application to arithmetic complexity and statistics. Linear Algebra Appl. 1977;18(2):95–138. 10.1016/00243795(77)900696.Search in Google Scholar
[35] Boyd S, Parikh N, Chu E, Peleato B, Eckstein J. Distributed optimization and statistical learning via the alternating direction method of multipliers found. Trends Mach Learn. Jan. 2011;3(1):1–122.10.1561/9781601984616Search in Google Scholar
[36] Iserles A, Golub GH, Van Loan CF. Matrix computations. Vol. 74. Issue 469. Baltimore, MD: Johns Hopkins University Press; 1990.Search in Google Scholar
[37] Glowinski R, Tallec PL. Augmented lagrangian and operator splitting methods in nonlinear mechanics. Philadelphia, PA, USA: SIAM; 1989.10.1137/1.9781611970838Search in Google Scholar
[38] AhmadiAsl S, Abukhovich S, AsanteMensah MG, Cichocki A, Phany AH, Tanaka T. Randomized algorithms for computation of tucker decomposition and higher order SVD (HOSVD). IEEE Access. 2021;9:28684–706. 10.1109/access.2021.3058103.Search in Google Scholar
[39] Manjón JV, Coupé P, MartíBonmatí L, Collins DL, Robles M. Adaptive nonlocal means denoising of MR images with spatially varying noise levels. J Magn Reson Imaging. 2010;31(1):192–203. 10.1002/jmri.22003.Search in Google Scholar PubMed
[40] Renard N, Bourennane S, BlancTalon J. Denoising and dimensionality reduction using multilinear tools for hyperspectral images. IEEE Geosci Remote Sens Lett. 2008;5(2):138–42. 10.1109/LGRS.2008.915736.Search in Google Scholar
[41] Nishimura Y, Suzuki T, Fukuda K, Fukuta M. Study of image reconstruction by UT probe array using truncated singular value decomposition. Int J Appl Electromagn Mech. 2014;45(1–4):21–6. 10.3233/JAE141808.Search in Google Scholar
[42] Zhang L, Zhang L, Mou X, Zhang D. FSIM: A feature similarity index for image quality assessment. IEEE Trans Image Process. 2011;20(8):2378–86. 10.1109/TIP.2011.2109730.Search in Google Scholar PubMed
[43] Wald L. Data fusion. Definitions and architectures – Fusion of images of different spatial resolutions. Vol. 8. Paris, France: Presses de l’Ecole, Ecole des Mines de Paris; 2002.Search in Google Scholar
© 2022 Hao Wu et al., published by De Gruyter
This work is licensed under the Creative Commons Attribution 4.0 International License.