검색 전체 메뉴
PDF
맨 위로
OA 학술지
Simple Denoising Method for Novel Speckle-shifting Ghost Imaging with Connected-region Labeling
  • CC BY-NC
  • CC BY-NC
ABSTRACT

A novel speckle-shifting ghost imaging (SSGI) technique is proposed in this paper. This method can effectively extract the edge of an unknown object without achieving its clear ghost image beforehand. However, owing to the imaging mechanism of SSGI, the imaging result generally contains serious noise. To solve the problem, we further propose a simple and effective method to remove noise from the speckle-shifting ghost image with a connected-region labeling (CRL) algorithm. In this method, two ghost images of an object are first generated according to SSGI. A threshold and the CRL are then used to remove noise from the imaging results in turn. This method can retrieve a high-quality image of an object with fewer measurements. Numerical simulations are carried out to verify the feasibility and effectiveness.


KEYWORD
Ghost imaging , Edge detection , Connected-region labeling
  • I. INTRODUCTION

    Ghost imaging (GI), as a novel optical imaging technique, has been developed over the past two decades [1-7]. Different from classical optical imaging, GI requires two light beams: One is called the reference arm, which is recorded by a detector with spatial-resolution ability such as a charge-coupled device (CCD), while the other is the object arm, which is detected by a single-pixel detector (bucket detector) after the light interacts with the object to be measured. The object can be computationally imaged by correlating the detected data in the two arms. In 2008, Shapiro proposed a computational GI (CGI) technique that replaces the reference beam in the classical GI system with a virtual computational part [8]. Subsequently in 2009, Bromberg et al. designed an experimental system to verify the feasibility of CGI, in which a spatial light modulator (SLM) is used to generate a series of random intensity patterns as illuminations [9]. Owing to the imaging mechanism of CGI, the retrieved image generally contains serious noise, which has become an obstacle in some practical applications, although many efforts such as compressive GI [10], differential GI [11], normalized GI [12], and iterative GI [13, 14] have been made to improve the signal-to-noise ratio (SNR).

    Edge detection based on CGI has currently become a hot research topic. In 2015, Liu et al. first proposed an edge-detection method based on gradient GI (GGI) [15]. This method can extract the edge in a direction according to the gradient angle adopted in the random patterns in CGI. Based on GGI, Mao et al. recently proposed the speckle-shift GI (SSGI) to extract the edge without prior knowledge of an object [16]. These methods can bring great convenience in object recognition, but the noise is still difficult to remove, and will further affect the accuracy of recognition. To improve the quality of edge detection, Wang and Ren et al. have successively proposed subpixel-speckle-shifting ghost imaging [17] and difference Fourier single-pixel imaging [18], which can fully remove random noise from the imaging results, but the number of measurements should be double or even triple the resolution of the detected image.

    In this paper, we propose a novel SSGI method that can effectively extract the edge of an unknown object. In this method, the illumination intensity patterns can be structured in a computer beforehand, and only two groups of intensity patterns are used to project the object. Compared tto the existing SSGI method, the imaging speed and quality will be improved. However owing to the imaging mechanism of GI, the imaging result generally contains serious noise. To solve this problem, a simple and effective denoising method for SSGI with a connected-region labeling (CRL) algorithm is further explored. First, two edge images of an unknown object are extracted according to SSGI. A threshold and the CRL algorithm are then used to remove noise from the imaging results in turn. This method can realize the image of an object with high quality, using fewer measurements. The organization of this paper is as follows. In Section II, the principle of the novel SSGI with the Sobel operator is introduced. In Section III, the steps of the denoising method with CRL are described specifically. Numerical simulation results are presented in Section IV, and finally we conclude this paper in Section V.

    II. A NOVEL SSGI WITH THE SOBEL OPERATOR

    In Ref. [16], the edge-detection methods using SSGI with the Laplacian operator, Standard operator, and Sobel operator have been demonstrated theoretically and experimentally. The results show that SSGI with the Sobel operator is robust in noisy environments. However, the intensity patterns used to illuminate the object are divided into 8 groups. Only one eighth of the intensity patterns are used to detect the edge in a corresponding direction, so the clarity of the imaging result will be poor if the number of measurements is insufficient. Here we introduce a novel SSGI technique with the Sobel operator, in which the intensity patterns are divided into two groups, i.e.

    image

    and

    image

    where and are the horizontal and vertical gradient patterns computed from an arbitrarily selected random intensity pattern Ir(x, y), r = 1, 2, ··· M, and M is the number of measurements in the SSGI. All of these intensity patterns are computed and stored in the computer beforehand.

    The setup for SSGI is schematically shown in Fig. 1. In the system, the series of structured patterns is generated by the computer-controlled digital micromirror device (DMD) and projected onto the object O(x, y). The modulated light is finally collected by a single-pixel (bucket) detector (BD) to form the measurement data, i.e. {} and {},

    image

    and

    image

    The horizontal and vertical edge of the object are imaged by correlating the detected data with the random intensity patterns, i.e.,

    image

    and

    image

    where <∙> denotes the mean of the elements. The edge of the object is obtained by

    image

    Further explanation of the edge-imaging process is presented in the appendix.

    Compared to conventional SSGI [16], this method has two advantages. One is that the imaging speed will be increased, because the horizontal and vertical intensity patterns can be computed beforehand, and only two groups intensity patterns are used to illuminate the object; in conventional SSGI, eight groups of intensity patterns need to be generated by the DMD and projected onto the object. The other advantage is that the imaging result may be clearer than that from SSGI with the same number of measurements. For example, 20,000 measurements are adopted in conventional SSGI, but only 2500 intensity patterns are used to extract the edge in one direction. In our method, the edge in the horizontal or vertical direction will be imaged by 10,000 measurements. Here, a binary image with 128 × 128 pixels (Fig. 2(a)) is adopted to verify the method’s performance by numerical simulations. For comparison, the edge extracted from the original image by Sobel operator is shown in Fig. 2(b). One of the intensity patterns in the horizontal and vertical directions computed from Eqs. (1) and (2) is presented in Figs. 2(c) and 2(d) respectively. The horizontal and vertical edges of the object are imaged and shown in Figs. 2(e) and 2(f) respectively. Combining the two edge images according to Eq. (7), the result is shown in Fig. 2(g). As can be seen from the results, the edge imaged with our method is clearer than that from the conventional SSGI (Fig. 2(h)).

    III. DENOISING METHOD

    Owing to the imaging mechanism of the speckle-pattern illumination, no matter for our method or the conventional SSGI, the detected results (Figs. 2(g) and 2(h)) generally contain serious noise, which may be an obstacle in some potential applications. Considering that the speckle noise is randomly distributed in the SSGI result, but the detected edge of the object is nearly fixed and constant, we therfore intend to reduce noise by comparing two SSGI results E1(x, y) and E2(x, y). The flow chart for the denoising method is shown in Fig. 3. According to the edge-detection results, an appropriate threshold is first set as

    image

    where μ and σ respectively denote the mean and standard deviation of E(x, y). If E1(x, y) ≥ T1 and E2(x, y) ≥ T2, the position (x, y) is believed to be the detected edge; otherwise, it is speckle noise. The preprocessing result can be mathematically expressed as

    image

    where Esum(x, y) is the sum of the two SSGI results E1(x, y) and E2(x, y). Then the CRL, as a famous morphology algorithm [19, 20], is also used here to further reduce the small-speckle noise. We define the small-speckle area as s < εsmax. The noise-reduction result using CRL can be written as

    image

    where smax is the largest connected region and ε is a coefficient. In this method, the sparse matrix and Dulmage-Mendelsohn decomposition algorithm are used to label the largest connected region [21].

    IV. NUMERICAL SIMULATIONS

    In this section, we will verify the feasibility of our proposed method by numerical simulations. To evaluate the quality of the detected edge, in this paper the peak signal-to-noise ratio (PSNR) is used as a metric, and defined as

    image

    where g′(x, y) and g(x, y) denote the detected result and reference image respectively. M × N is the size of the image, and (2k − 1) represents the maximum grayscale value.

    Apart from the binary image shown in Fig. 2(a), the grayscale image with 128 × 128 pixels shown in Fig. 4(a) is also adopted as the original object image in this simulation. The edge shown in Fig. 4(b) is directly extracted with the Sobel operator for comparison.

    In the following simulations, 20,000 structured patterns are divided into two groups (i.e. 10,000 patterns in each group) to illuminate the object image, and each group will be used to generate a ghost image. Correlating the detected data with the structured patterns, the two edge images are generated by Eqs. (5)~(7) (i.e. the novel SSGI with the Sobel operator), and shown in Figs. 5(a), 5(b), 5(f) and 5(g) with PSNR = −2.18, −2.21, −6.11, and −6.09 dB respectively. As can be seen from the results, the edge looms through the speckle noise for both the binary and grayscale images. Because the noise is randomly distributed in the edge-detection results with the same probability, it will be tend to uniform in a superposition of the two results, which can be seen in Figs. 5(c) and 5(h). Subsequently using the threshold, the preprocessing results are obtained by Eq. (9) and shown in Figs. 5(d) and 5(i), respectively. Although most of the noise has been removed, some scattered noise still remains in the results. To further remove the small-speckle noise, the CRL algorithm is used here according to Eq. (10) for ε = 0.01, and the results are shown in Figs. 5(e) and 5(j) with PSNR = 19.54 and 8.48 dB, corresponding to the binary and grayscale image respectively.

    Apart from binary and grayscale images, the method proposed in this paper is also suitable for color images, because a color image can be decomposed into red (R), green (G), and blue (B) components. The connected region is labeled in a binary image, so the image retrieved with SSGI should be binarized at first, and the final extracted edge is binary. Figures 6(a) and 6(b) respectively present a color image and its extracted edge with this proposed method. We can see that the result is similar to that for the grayscale image.

    Recently, Wang et al. have proposed an edge-detection method based on subpixel-shifting ghost imaging (SPSGI) [17]. According to the orthogonality of the Walsh-Hadamard matrix, the edge of an unknown object can be obtained with high quality and resolution by correlating the detected data with the subpixel-shifted Walsh-Hadamard patterns. In addition, Ren et al. have also proposed another edge-detection method with difference Fourier single-pixel imaging (DFSI) [18]. These two methods can extract the edge of an object with high quality, but the number of measurements should be at least double the resolution of the detected image for SPSGI, and triple for DFSI. Here we also take Figs. 2(a) and 4(a) as examples to numerically simulate the SPSGI and DFSI techniques.

    When 32,768 Walsh-Hadamard patterns and 49,152 sinusoidal patterns are respectively used in SPSGI and DFSI, the edge is extracted clearly for both binary and grayscale images, as seen in Figs. 7(a)~7(d). However, when the number of measurements is insufficient, the imaging results blur. To intuitively observe the differences among the imaging results of these methods, 30,000 structured patterns are adopted to detect the edge of the binary image, and the results are shown on the right side of Fig. 8. As can be seen, the result is overlapped for SPSGI and blurred for DFSI. Here we only take the binary image as an example and test the PSNRs of the imaging results, varying with the number of measurements, for SSGI, DFSI, SPSGI, and our method proposed in this paper. The curves are shown in Fig. 8. As can be seen from the results, our proposed method can realize high imaging quality when fewer measurements are adopted, although it cannot completely remove noise from the detected edge.

    V. CONCLUSION

    In this paper we have proposed a novel SSGI technique to improve imaging speed and quality. Numerical simulations are carried out to verify its performance. However, owing to the imaging mechanism of SSGI, the results still contain serious noise. To solve the problem, we also explored a method to suppress the noise of SSGI results via CRL algorithms. In this method, two frame-edge-detection results are first generated by SSGI with the Sobel operator, and then a preprocessing step with a threshold and the CRL algorithm are applied in turn to reduce the noise. Compared to SPSGI and DFSI techniques, this method cannot completely remove noise from the detected edge, but it can realize high imaging quality with fewer measurements. Numerical simulations are carried out to verify the method’s feasibility and effectiveness. Actually, the method proposed in this paper is simple, and so will take little operation time. The disadvantage of this method is that the CRL is invalid for noise adhering to the edge; some other effective algorithms still need to be further investigated.

참고문헌
  • 1. Pittman T. B., Shih Y. H., Strekalov D. V., Sergienko A. V. 1995 Optical imaging by means of two-photon quantum entanglement [Phys. Rev. A] Vol.52 P.R3429-R3432 google cross ref
  • 2. Bennink R. S., Bentley S. J., Boyd R. W. 2002 “Two-photon” coincidence imaging with a classical source [Phys. Rev. Lett.] Vol.89 P.113601 google cross ref
  • 3. Valencia A., Scarcelli G., Angelo M. D., Shih Y. 2005 Two-photon imaging with thermal light [Phys. Rev. Lett.] Vol.94 P.063601 google cross ref
  • 4. Liu X. F., Chen X. H., Yao X. R., Yu W. K., Zhai G. J., Wu L. A. 2014 Lensless ghost imaging with sunlight [Opt. Lett.] Vol.39 P.2314-2317 google cross ref
  • 5. Zhang Z., Ma X., Zhong J. 2015 Single-pixel imaging by means of Fourier spectrum acquisition [Nat. Commun.] Vol.6 P.6225-6230 google cross ref
  • 6. Wang L., Zhao S. M. 2016 Fast reconstructed and high-quality ghost imaging with fast Walsh-Hadamard transform [Photonics Res.] Vol.4 P.240-244 google cross ref
  • 7. Jue W., Renlong Y., Yu X., Yanming S., Yanru C., Qi Z. 2016 Ghost imaging with different speckle sizes of thermal light [J. Opt. Soc. Korea] Vol.20 P.8-12 google cross ref
  • 8. Shapiro J. H. 2008 Computational ghost imaging [Phys. Rev. A] Vol.78 P.061802 google cross ref
  • 9. Bromberg Y., Katz O., Silberberg Y. 2009 Ghost imaging with a single detector [Phys. Rev. A] Vol.79 P.053840 google cross ref
  • 10. Katz O., Bromberg Y., Silberberg Y. 2009 Compressive ghost imaging [Appl. Phys. Lett.] Vol.95 P.131110 google cross ref
  • 11. Ferri F., Magatti D., Lugiato L. A., Gatti A. 2010 Differential ghost imaging [Phys. Rev. Lett.] Vol.104 P.253603 google cross ref
  • 12. Sun B., Welsh S. S., Edgar M. P., Shapiro J. H., Padgett M. J. 2012 Normalized ghost imaging [Opt. Express] Vol.20 P.16892-16901 google cross ref
  • 13. Wang W., Wang Y. P., Li J., Yang X., Wu Y. 2014 Iterative ghost imaging [Opt. Lett.] Vol.39 P.5150-5153 google cross ref
  • 14. Wang W., Hu X., Liu J., Zhang S., Suo J., Situ G. 2015 Gerchberg-Saxton-like ghost imaging [Opt. Express] Vol.23 P.28416-28422 google cross ref
  • 15. Liu X. F., Yao X. R., Lan R. M., Wang C., Zhai G. J. 2015 Edge detection based on gradient ghost imaging [Opt. Express] Vol.23 P.33802-33811 google cross ref
  • 16. Mao T., Chen Q., He W., Zou Y., Dai H., Gu G. 2016 Speckle-Shifting ghost imaging [IEEE Photonics J.] Vol.8 P.1-10 google
  • 17. Wang L., Zou L., Zhao S. 2018 Edge detection based on subpixel-speckle-shifting ghost imaging [Opt. Commun.] Vol.407 P.181-185 google cross ref
  • 18. Ren H., Zhao S., Gruska J. 2018 Edge detection based on single-pixel imaging [Opt. Express] Vol.26 P.5501-5511 google cross ref
  • 19. Jiang F., Zhang S., Wu S., Gao Y., Zhao D. 2015 Multi-layered gesture recognition with kinect [J. Mach. Learn. Res.] Vol.16 P.227-254 google
  • 20. Chen X., Wang Y., Wang Y., Ma M., Zeng C. 2016 Quantized phase coding and connected region labeling for absolute phase retrieval [Opt. Express] Vol.24 P.28613-28624 google cross ref
  • 21. Dillencourt M. B., Samet H., Tamminen M. 1992 A general approach to connected-component labeling for arbitrary image representations [J. ACM] Vol.39 P.253-280 google cross ref
이미지 / 테이블
  • [ ] 
  • [ ] 
  • [ FIG. 1. ]  Schematic diagram of the edge-detection system based on SSGI.
    Schematic diagram of the edge-detection system based on SSGI.
  • [ ] 
  • [ ] 
  • [ ] 
  • [ ] 
  • [ ] 
  • [ FIG. 2. ]  Numerical simulation results for the novel and conventional SSGI techniques: (a) original image, (b) edge directly extracted from the object with the Sobel operator, (c) and (d) the horizontal and vertical intensity patterns, (e) and (f) the extracted horizontal and vertical edges, (g) and (h) the imaging results for the novel and conventional SSGI respectively.
    Numerical simulation results for the novel and conventional SSGI techniques: (a) original image, (b) edge directly extracted from the object with the Sobel operator, (c) and (d) the horizontal and vertical intensity patterns, (e) and (f) the extracted horizontal and vertical edges, (g) and (h) the imaging results for the novel and conventional SSGI respectively.
  • [ FIG. 3. ]  Flow chart for the denoising method.
    Flow chart for the denoising method.
  • [ ] 
  • [ ] 
  • [ ] 
  • [ ] 
  • [ FIG. 4. ]  (a) The other object image, and (b) its edge, directly extracted with the Sobel operator.
    (a) The other object image, and (b) its edge, directly extracted with the Sobel operator.
  • [ FIG. 5. ]  Numerical simulation results for our proposed method in this paper, for number of measurements M = 20000 and coefficient ε = 0.01 in CRL.
    Numerical simulation results for our proposed method in this paper, for number of measurements M = 20000 and coefficient ε = 0.01 in CRL.
  • [ FIG. 6. ]  Simulation result for a color image: (a) original image, (b) edge extracted with the proposed method.
    Simulation result for a color image: (a) original image, (b) edge extracted with the proposed method.
  • [ FIG. 7. ]  Numerical simulation results for ((a) and (b)) SPSGI and ((c) and (d)) DFSI.
    Numerical simulation results for ((a) and (b)) SPSGI and ((c) and (d)) DFSI.
  • [ FIG. 8. ]  The curves for PSNR varying with the number of measurements in SSGI, DFSI, SPSGI, and our proposed method, for the binary image.
    The curves for PSNR varying with the number of measurements in SSGI, DFSI, SPSGI, and our proposed method, for the binary image.
(우)06579 서울시 서초구 반포대로 201(반포동)
Tel. 02-537-6389 | Fax. 02-590-0571 | 문의 : oak2014@korea.kr
Copyright(c) National Library of Korea. All rights reserved.