Table of Contents Author Guidelines Submit a Manuscript
Advances in Astronomy
Volume 2018, Article ID 2417939, 7 pages
https://doi.org/10.1155/2018/2417939
Research Article

Poisson Denoising for Astronomical Images

1Military College of Signals, National University of Sciences and Technology, Pakistan
2Center for Advanced Studies in Telecommunication, COMSATS Islamabad, Pakistan

Correspondence should be addressed to Abdul Ghafoor; kp.ude.tsun@scm-roofahgludba

Received 12 February 2018; Revised 5 April 2018; Accepted 23 April 2018; Published 10 June 2018

Academic Editor: Wing-Huen Ip

Copyright © 2018 Fahad Shamshad et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

A denoising scheme for astronomical color images/videos corrupted with Poisson noise is proposed. The scheme employs the concept of Exponential Principal Component Analysis and sparsity of image patches. The color space RGB is converted to YCbCr and -means++ clustering is applied on luminance component only. The cluster centers are used for chromatic components to improve the computational efficiency. For videos, the information of both spatial and temporal correlations improves the denoising. Simulation results verify the significance of proposed scheme in both visual and quantitative manner.

1. Introduction

Recent advancements in astronomy and digital systems emphasize the development of more sophisticated image processing algorithm. Acquisition of astronomical images in low photon count region is a dominant source of Poisson noise. The images containing intensity dependent Poisson noise (having same mean and variance value) cannot be accurately modelled with Gaussian distribution.

In [1], Poisson noise is converted into Gaussian noise with unit variance using Variance Stabilization Transform (VST). Further improvement is made by applying Bayesian multiscale likelihood models [2] and penalization hypothesis testing [3]. However these schemes [13] provide significantly degraded results for high Poisson noise. Close form approximation for unbiased inverse of Anscombe transform yields increased error peak for Poisson noise [4, 5]. Structured dictionary learning approach exploiting the self-similarity of image patches [6] requires significant amount of computational time especially for color images and videos.

In [7], multiscale Poisson denoising based on shrinkage operators to attain maximum error is proposed. However distance is not suitable for Poisson corrupted measurements. In [8] spatially adaptive total variation framework is proposed with split Bregman optimization for Poisson denoising to handle high computational load but yields poor results in low photon count regime which is usually the case with astronomical images. Directional lapped orthogonal transform overcomes the limitation of diagonal edges and textures which separable wavelets fail to denoise efficiently [9]. Computationally efficient technique for estimation of unknown noise parameters requires a blank image under same mechanical shutter [10]. In a nutshell, most of above-mentioned techniques either suffer from high computational complexity or give poor results in high Poisson noise.

A denoising scheme for astronomical color images/videos corrupted with Poisson noise is proposed. The scheme first converts the color space from RGB to YCbCr and applies -means++ clustering on luminance component only. The same cluster centers are used for chromatic components to improve the computational efficiency. Framework of Exponential Principal Component Analysis (EPCA) [11] is employed that better suits high Poisson noise. The proposed algorithm addresses several drawbacks of existing schemes while ingeriting their main strengths. For instance, patch based approach is used to exploit redundancies in image that have been shown to improve performance of image denoising algorithms [6]. Similarly clustering algorithms are used in different denoising algorithms to group similar patches together [5]. Note that careful clustering has significant importance in case of high Poisson noise (where only few photons are acquired by detector). Similarly taking into account temporal correlation of videos also improves the performance of restoration algorithms [12, 13]. Simulation results on different images and videos verify the significance of proposed scheme in both visual and quantitative manner.

2. Proposed Technique

Let be the observed image from image acquisition device having dimensions , where are rows, are columns, and green, are color bands. For let (column stack representation of ) be the observed pixel at location whose entries are independent Poisson random variables with true value which is to be estimated. Given the clean image likelihood of observing noisy Poisson image can be written asNoisy image is converted into luminance chrominance (YCbCr) color space by matrix transformation asLet denote the noisy image in YCbCr color space where ; then matrix of vectorized patches is constructed from the component by extracting all overlapping patches of size aswhere represents pixel in patch for and (considering image border issues) and let be defined for clean image as for noisy one. Similarly matrices and are built for and , respectively.

-means++ algorithm (having more accuracy and efficiency as compared to conventional -means) is then applied on matrix to obtain clusters and denote for the size of th cluster. As (luminance component) contains most of the valuable information (edges, shades, texture, etc.) of image the same cluster centers are then assigned to and components for computational efficiency of algorithm.

Different algorithms have used PCA to project the extracted noisy image patches to low dimensional space. Let be the pixel of patch in cluster ; using PCA it can be written aswhere is coefficient matrix of size and is dictionary atom matrix of size . is cluster size (patches in cluster ) and is number of principal axes to be retained. However, in this paper, the framework of [11] is used as it is more suitable for exponentially distributed data. Equation (4) can be written aswhere the exponential operator is element-wise.

After neglecting the constant term and applying maximization of negative Poisson likelihood alongwith approximation of (5), the loss function for cluster isEquation (6) is solved using fast Newton method [14], where and are initialized randomly. Then rows of (denoted as ) at iteration ) for cluster are updated aswhere , is diagonal matrix of size for cluster and is omitted for simplicity. Similarly columns of (denoted as ) are updated aswhere , .

Row of and columns of are updated iteratively. Stopping criteria for proposed algorithm are either number of iterations (maximum 20) or a defined error threshold (0.01), whichever is reached first. Equation (6) after solving reduces toand denoising estimate for cluster is ( is clean original image while is denoised image)After denoising of all clusters, denoised patches are projected onto pixels. As each pixel has multiple presentation in different patches, it is done by uniformly averaging all estimates of those patches containing given pixel. Denoising image is then converted to RGB domain as

The above denoising methodology is extended for Poisson noise corrupted astronomical videos by taking spatial correlation into account. Let be Poison corrupted frame of size in YCbCr color space where . If all overlapping cubes of size are extracted then total number of cubes will be , where represents number of frames. For simplicity denote by and by (total cubes then become ). After concatenating every cube row-wise, patch by patch, in single row matrix form , pixel of cube in frame is written as

Note that index is changed after every pixel as these are total entries on front side of cube (size ). -means++ clustering is then applied on these cubes and clusters are denoised using (6) (cubes are vectorized in this case). After denoising these cubes are projected onto pixels by averaging and converted back to RGB color space.

3. Results and Discussion

The proposed and state-of-the-art existing Poisson denoising techniques are applied on various astronomical images/videos (taken using Hubble telescope). For images, the proposed technique is compared with K-SVD [15], CBM3D [16], PLOW [17], Anscombe transform [18], and NLPCA [5]. The results on videos are compared with KSVD-3D [12], CBM4D [13], and PURE-LET [19].

Figures 13 show visual comparison of existing and proposed techniques on three astronomical images (Disk, Nebula, and Jupiter) with different noise peaks. It can be observed that the proposed technique is able to reconstruct a better output image as compared to existing techniques (even in low noise peak values). Tables 1 and 2 provide quantitative analysis of existing and proposed techniques using Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) [20]. Note that the proposed technique provides better quantitative values as compared to the state-of-the-art existing techniques.

Table 1: PSNR comparison.
Table 2: SSIM comparison.
Figure 1: Disk: noisy input, reference, and denoised images (for peak = 0.2).
Figure 2: Nebula: noisy input, reference, and denoised images (for peak = 0.2).
Figure 3: Jupiter: noisy input, reference, and denoised images (for peak = 1).

Figure 4 shows denoising result of proposed and existing techniques with cube size and noise peak = 1. The reconstructed frames (in Figures 4(i) and 4(j)) using proposed technique are quite similar to the reference frames (in Figures 4(k) and 4(l)). Figure 5 shows the PSNR and SSIM values of individual frames for different noise peaks. It can be observed that the proposed technique provides better quantitative values almost for all frames as compared to the state-of-the-art existing techniques. Table 3 shows the average quantitative results (PSNR and SSIM) of video frames (311–325). For Figure 5, the Saturn video created from Hubble images taken over about 9-hour span is used. Note that the PSNR and SSIM of only 10 frames (from 315 to 325) are shown in the paper, as scenes in video do not change much due to slow motion of Saturn. Consequently, the PSNR and SSIM values of frames are almost constant.

Table 3: Average PSNR (dB) and SSIM values of 311–325 frames.
Figure 4: Frames 311 and 318 of a Saturn video.
Figure 5: (a)–(c) PSNR graphs for peaks 0.2, 0.6, and 1; (d)–(f) SSIM graphs for peaks 0.2, 0.6, and 1.

4. Conclusion

A Poisson denoising scheme in nonlocal framework using EPCA based on the approach of Gaussian mixture models is proposed for astronomical imaging. The images patches sparsity and dictionary learning approach are unified. The method employs EPCA (which is more suitable for Poisson noise). The scheme first converts the color space from RGB to YCbCr and applies -means++ clustering on of luminance component only. The same cluster centers are used for chromatic components to improve the computational efficiency. The cubes (in case of color videos) of luminance components (information of both spatial and temporal correlations) are used for improved denoising. Simulation results verify the significance of proposed scheme in both visual and quantitative manner.

Data Availability

The data (images/videos) used in the paper can be provided on demand.

Conflicts of Interest

The authors declare no conflicts of interest regarding publication of this article.

References

  1. P. Fryzlewicz and G. P. Nason, “A Haar-Fisz algorithm for Poisson intensity estimation,” Journal of Computational and Graphical Statistics, vol. 13, no. 3, pp. 621–638, 2004. View at Publisher · View at Google Scholar · View at MathSciNet
  2. M. Jansen, “Multiscale Poisson data smoothing,” Journal of the Royal Statistical Society: Series B (Statistical Methodology), vol. 68, no. 1, pp. 27–48, 2006. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  3. B. Zhang, J. M. Fadili, and J.-L. Starck, “Wavelets, ridgelets, and curvelets for Poisson noise removal,” IEEE Transactions on Image Processing, vol. 17, no. 7, pp. 1093–1108, 2008. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  4. M. Makitalo and A. Foi, “A closed-form approximation of the exact unbiased inverse of the Anscombe variance-stabilizing transformation,” IEEE Transactions on Image Processing, vol. 20, no. 9, pp. 2697-2698, 2011. View at Publisher · View at Google Scholar · View at MathSciNet
  5. J. Salmon, Z. Harmany, C.-A. Deledalle, and R. Willett, “Poisson noise reduction with non-local PCA,” Journal of Mathematical Imaging and Vision, vol. 48, no. 2, pp. 279–294, 2014. View at Publisher · View at Google Scholar · View at MathSciNet
  6. R. Giryes and M. Elad, “Sparsity based Poisson denoising,” in Proceedings of the 2012 IEEE 27th Convention of Electrical and Electronics Engineers in Israel, IEEEI 2012, isr, November 2012. View at Publisher · View at Google Scholar · View at Scopus
  7. W. Cheng and K. Hirakawa, “Minimum risk wavelet shrinkage operator for Poisson image denoising,” IEEE Transactions on Image Processing, vol. 24, no. 5, pp. 1660–1671, 2015. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  8. A. Mansouri, F. Deger, M. Pedersen, J. Y. Hardeberg, and Y. Voisin, “An adaptive spatial–spectral total variation approach for Poisson noise removal in hyperspectral images,” Signal, Image and Video Processing, vol. 10, no. 3, pp. 447–454, 2016. View at Publisher · View at Google Scholar · View at Scopus
  9. Z. Chen and S. Muramatsu, “Poisson denoising with multiple directional lots,” in Proceedings of the ICASSP 2014 - 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1225–1229, Florence, Italy, May 2014. View at Publisher · View at Google Scholar
  10. X. Jin, Z. Xu, and K. Hirakawa, “Noise parameter estimation for Poisson corrupted images using variance stabilization transforms,” IEEE Transactions on Image Processing, vol. 23, no. 3, pp. 1329–1339, 2014. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  11. M. Collins, S. Dasgupta, and R. E. Schapire, “A generalization of principal component analysis to the exponential family,” in Proceedings of the 15th Annual Neural Information Processing Systems Conference, NIPS 2001, pp. 617–624, December 2001. View at Scopus
  12. M. Protter and M. Elad, “Image sequence denoising via sparse and redundant representations,” IEEE Transactions on Image Processing, vol. 18, no. 1, pp. 27–35, 2009. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  13. M. Maggioni, G. Boracchi, A. Foi, and K. Egiazarian, “Video denoising, deblocking, and enhancement through separable 4-D nonlocal spatiotemporal transforms,” IEEE Transactions on Image Processing, vol. 21, no. 9, pp. 3952–3966, 2012. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  14. G. J. Gordon, “Generalized 2 linear 2 models,” Advances in Neural Information Processing Systems, pp. 577–584, 2002. View at Google Scholar
  15. L. N. Smith and M. Elad, “Improving dictionary learning: Multiple dictionary updates and coefficient reuse,” IEEE Signal Processing Letters, vol. 20, no. 1, pp. 79–82, 2013. View at Publisher · View at Google Scholar · View at Scopus
  16. K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Color image denoising via sparse 3D collaborative filtering with grouping constraint in luminance-chrominance space,” in Proceedings of the IEEE International Conference on Image Processing, pp. 313–316, San Ontonio, Tex, USA, 2007.
  17. P. Chatterjee and P. Milanfar, “Patch-based near-optimal image denoising,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1635–1649, 2012. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  18. M. Makitalo and A. Foi, “Optimal inversion of the Anscombe transformation in low-count Poisson image denoising,” IEEE Transactions on Image Processing, vol. 20, no. 1, pp. 99–109, 2011. View at Publisher · View at Google Scholar · View at MathSciNet
  19. F. Luisier, C. Vonesch, T. Blu, and M. Unser, “Fast interscale wavelet denoising of Poisson-corrupted images,” Signal Processing, vol. 90, no. 2, pp. 415–427, 2010. View at Publisher · View at Google Scholar · View at Scopus
  20. Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. View at Publisher · View at Google Scholar · View at Scopus