7 May 2014 Regularization approach to scene-based nonuniformity correction
Author Affiliations +
Abstract
Various scene-based nonuniformity correction (SBNUC) methods have been proposed to diminish the residual nonuniformity (RNU) of the infrared focal plane array (IRFPA) sensors. Most existing SBNUC techniques require a relatively large number of image frames to reduce the RNU. In some applications, however, there is not enough time for capturing a large number of image frames prior to the camera operation, or only several image frames are available to users. A new scene-based approach that can correct the RNU using only several image frames is proposed. The proposed method formulates the SBNUC process as an energy minimization problem. In the proposed energy function, we introduce regularization terms for the parameter regarding the responsivity of the IRFPA as well as for the true scene irradiance. Correction results are obtained by minimizing the energy function using a numerical technique. Experimental results demonstrate the effectiveness of the proposed method.
Kim, Kim, Kim, Lee, and Lee: Regularization approach to scene-based nonuniformity correction

1.

Introduction

The spatial noise caused by nonuniformity of individual detector elements in the infrared focal plane array (IRFPA) can limit the overall performance of imaging systems. In general, the responsivity of individual detector elements is assumed to be linear. Therefore, it is possible to correct the nonuniformity using the well-known two-point correction (TPC) method, where two blackbodies at different temperatures are employed.1 Since the nonuniformity normally drifts in time2 and the correction capability of the TPC is degraded with repeated operations,3 we need to recalibrate the IRFPA. In practice, there are two difficulties in using the TPC method to compensate the residual nonuniformity (RNU) resulting from the temporal drifit or repeated operations: (1) we have to maintain two distinct heat sources in imaging systems; and (2) real-time video operation is interrupted during the correction process.4

Various scene-based nonuniformity correction (SBNUC) algorithms have been proposed to solve these problems. In general, SBNUC schemes can be broadly divided into two categories: constant statistics (CS) methods45.6 and least mean square (LMS) methods.78.9.10 The original CS method assumes that the temporal mean and standard deviation of each pixel are constant over time and space.5 The performance of the original CS method is reliable as long as the assumption is valid. As pointed out in Refs. 4 and 6, however, thousands of image frames are required to hold the assumption. Zhang and Zhao4 proposed a local constant statistics (LCS) method, which assumes that the temporal statistics are constant in a local region around each pixel. The LCS method improved the correction performance for the same number of input frames.4 Later, Zuo et al.6 generalized the LCS method by introducing a new constraint called multiscale CS.

In the LMS methods, the correction parameters are learned by minimizing the LMS error between corrected images and desired output images. The minimization is performed frame-by-frame using the stochastic gradient descent (SGD) technique.11 Since the desired output images are not available, spatially low-pass filtered input images are used as the desired ones.7 However, the performance of the LMS method is degraded at strong edge points as reported in Refs. 8 to 10. Thus, several methods improve the estimation accuracy of correction parameters by suppressing the influence of strong edge points in the minimization process. Vera and Torres8 adaptively adjust the learning rate which is a fixed parameter in the original SGD techinque, according to the local spatial standard deviations of input images. In this way, the influence of strong edge regions, where the local standard deviation is large, is reduced in the minimization process. This method is further improved in Ref. 9 to remove burn-in artifacts caused by temporally slowly-varying image regions. An approach proposed in Ref. 10 updates correction parameters only when sufficient change occurs between consecutive images.

Lately, different error functions are proposed as alternatives to the LMS error in the original LMS method.1213.14 Interframe registration-based LMS method12,13 first registers the previous corrected image with the current input image by assuming that only slight translational motion exists. Then, this technique minimizes the LMS error between the previously shifted image and the currently corrected image. Vera et al.14 minimize the total variation of corrected images to obtain the correction parameters.

Although these previous SBNUC algorithms may reduce the RNU, they have a common problem: a relatively large number of image frames are required to acquire the correction parameters.15 Two recently proposed SBNUC algorithms perform NUC using several image frames16 or even only two image frames.15 These methods can achieve good performance when the relative motion in successive image frames is a small translation along vertical or horizontal directions. However, large displacements between consecutive image frames can occur in some applications.17,18 To deal with the motion constraint of the previous approaches,15,16 we propose a new SBNUC method that estimates the correction parameters using several image frames. In the proposed method, we utilize the prior information on the parameters regarding the responsivity and the true scene irradiance. There is no restriction on the motion in successive image frames unless it is static.

The rest of this article is organized as follows. In Sec. 2, the proposed SBNUC method is detailed. In Sec. 3, the performance of the proposed method is evaluated. Conclusion is drawn in Sec. 4.

2.

Proposed Method

In this section, we first formulate an optimization problem for correcting the RNU followed by its numerical solution.

2.1.

Formulation

Let us assume that the characteristic of each detector element in the IRFPA is linear.5,7 Then, the acquired signal y(i,j,t) for the (i,j)’th detector element at time t is given by

(1)

y(i,j,t)=x(i,j,t)+o(i,j,t),
where o(i,j,t) represents the offset of each detector element and x(i,j,t) indicates the scene irradiance. Here, we assume that there is no gain nonuniformity since the offset component is the dominant source of the RNU.15,19,20 Given the image observation model [Eq. (1)], our objective is to estimate x(i,j,t) and o(i,j,t). This problem can be solved by minimizing the proposed energy function, which consists of three terms:

(2)

E(x,o)=tΩi,jID(x,o)+λoC1(o)+λxC2(x),
where Ω and I denote the set of given image frames and the image domain, respectively. λo and λx are regularization parameters for C1(o) and C2(x), respectively. The data-fidelity term D which measures the mismatch between the observed image and the estimates is given by

(3)

D(x,b)=[y(i,j,t)x(i,j,t)o(i,j,t)]2.

Solving Eq. (3) alone is an underconstrained problem in which the number of unknowns is greater than that of equations. Thus, a regularization approach is taken to estimate x(i,j,t) and o(i,j,t) in this work. The regularization term for the offset o(i,j,t) is defined as follows:

(4)

C1(o)=[o(i,j,t)t]2.

C1(o) is derived from the observation that the offset changes very slowly in time.7,9 In other words, the offset remains almost constant for several consecutive image frames.20 This regularization term favors the offset with small changes along the time axis. If we correctly estimate o(i,j,t) for the given image frames, their temporal variation is negligible, which means that C1(o) is very close to zero.

The last term C2(x) in Eq. (2) is introduced to regularize the scene irradiance x(i,j,t). In general, x(i,j,t) is smooth in the spatial domain. This fact is implictly used in the original LMS method, where the desired image is the spatially low-pass filtered input image.7 Thus, it is natural for us to require x(i,j,t) to be smooth in the spatial domain. Since the degree of the spatial smoothness can be measured via the image gradient, C2(x) is given by

(5)

C2(x)=wi[y(i,j,t)][x(i,j,t)i]2+wj[y(i,j,t)][x(i,j,t)j]2.

The smoothness term C2(x) is proportional to the magnitude of the spatial intensity change of the scene irradiance. Therefore, the smoother the scene irradiance is, the smaller the value of C2(x) is. However, the smoothness constraint is not appropriate at edge points as pointed out in Sec. 1. Since the spatial variation of x(i,j,t) is normally greater than that of RNU,2 the large spatial variation in the input image y(i,j,t) is mainly due to the edge points of x(i,j,t). We adaptively adjust the effect of the smoothness constraint according to the gradient of y(i,j,t). The weighting factors wi[y(i,j,t)] and wj[y(i,j,t)] in Eq. (5) are defined as follows

(6)

wi(y)=[|y(i,j,t)i|γ+ε]1,

(7)

wj(y)=[|y(i,j,t)j|γ+ε]1.

The exponent γ controls the sensitivity to the spatial gradients of y(i,j,t) and ε is a small constant that prevents dividing by zero. Since the weighting factors are inversely proportional to the spatial gradients of y(i,j,t), the smoothness constraint has little effect on edge regions. These weighting factors are the same as smoothness weights for the image smoothing operator in Ref. 21.

2.2.

Numerical Solution

The proposed energy function E(x,o) in Eq. (2), a function of two variables, is nonconvex. We minimize the energy function by solving two convex subproblems in an alternating way with initial estimates x(0)=y and o(0)=0:

(8)

x(n)=argminxE[x,o(n1)],

(9)

o(n)=argminoE[x(n),o].

The above process is repeated until there is no significant change in the estimates x(n) and o(n).

We compute the partial derivative of Eq. (2) with respect to x in order to solve Eq. (8). First, we represent Eq. (2) in matrix notation as follows:

(10)

tΩ[y(t)x(t)o(t)]T[y(t)x(t)o(t)]+λx[x(t)TDiTWi(t)Dix(t)+x(t)TDjTWj(t)Djx(t)],
where y(t), x(t), and o(t) are the lexicographically ordered vectors corresponding to the acquired signal, the scene irradiance, and the offset at time t, respectively. Wi(t) and Wj(t) denote diagonal matrices containing values of the weighting factors wi(y) and wj(y) at time t, respectively. Di and Dj are the backward difference operators along i and j directions, which approximate the spatial partial gradients. Note that the regularization term for the offset C1(o) is omitted in Eq. (10) since C1(o) is constant with respect to x. Then, E/x in matrix notation is given by

(11)

Ex=tΩ[A(t)x(t)b(t)],
where

(12)

A(t)=λx[DiTWi(t)Di+DjTWj(t)Dj],

(13)

b(t)=y(t)o(t).

Therefore, we solve a large system of linear equations [i.e., A(t)x(t)=b(t)] for each x(t). The conjugate-gradient (CG) method is used to solve the linear equations in this work since the matrix A is sparse, symmetric, and positive definite.22

We rewrite Eq. (2) in different matrix notation from that represented in Eq. (10) to solve Eq. (9):

(14)

i,jI[y(i,j)x(i,j)o(i,j)]T[y(i,j)x(i,j)o(i,j)]+λo[o(i,j)TDtTDto(i,j)],
where y(i,j), x(i,j), and o(i,j) are vectors formed by lexicographically stacking the acquired signal, the scene irradiance, and the offset in the time domain, respectively, for each detector element located in (i,j). Dt is the temporal backward difference operator. Similar to Eq. (10), we exclude the regularization term C2(x) in Eq. (14) since it is constant with respect to o. Differentiation with respect to o produces

(15)

Eo=i,jI[Ao(i,j)b(i,j)],
where

(16)

A=I+λoDtTDt,

(17)

b(i,j)=y(i,j)x(i,j).

I denotes identity matrix in Eq. (16). The CG method is used here again to obtain the values of the offset for each o(i,j).

3.

Simulation Results

To our best knowledge, no study has been reported on correcting the RNU with several image frames that have large displacements. Therefore, no comparison is made with existing SBNUC methods in this work. The regularization parameters are empirically set to λo=6, λx=0.1 for all experiments in this article, and the value of γ is determined to be 0.2. At first, we test the convergence of the proposed method with eight synthetic images as shown in Fig. 1. The eight images are generated by adding the artificial offset9,15 to calibrated infrared images caputured by a 320×256 InSb focal plane array camera operating in the 3 to 5 μm range. The RNU is generally composed of two patterns, the low-frequency one and white noise-like one as reported in Ref. 23. However, only the white noise-like pattern is usually prominent to observers together with natural scenes. This is due to the masking effect of the human visual system, which attenuates contrast sensitivity at low-spatial frequencies.24 Therefore, the artificial offset is generated as realizations of independent identically distributed Gaussian random variables.9

Fig. 1

Simulated nonuniformity images.

OE_53_5_053105_f001.png

We plot the proposed energy function in Eq. (2) against the number of iterations. As shown in Fig. 2, the value of the energy function drops quickly. We obtain good results with 11 iterations in our expeirments. In Fig. 3, we present images corrected by the proposed NUC method. Close-up views of some parts of the images are depicted in Fig. 3(b) to help the reader observe the visual quality improvement. The proposed method produces acceptable results no matter how complex the spatial distribution is, as shown in Fig. 3.

Fig. 2

Convergence of the proposed method.

OE_53_5_053105_f002.png

Fig. 3

Nonuniformity correction (NUC) results on the synthetic images. (a) Whole images. (b) Close-ups of the input (left) and the proposed (right).

OE_53_5_053105_f003.png

We investigate the effect of the number of input images on the estimated scene irradiance using the eight synthetic images. Table 1 shows the peak signal-to-noise ratio results for different numbers of input images. As the number of input images increases, we can obtain more accurate results. This can be described by the fact that the information gained from consecutive image frames leads to high-quality NUC results and enhanced temporal consistency in the offset. Note that, however, raising the number of input images increases processing time. Thus, selecting a proper number of input images demands trade-off between the computational complexity and the image quality.

Table 1

Peak signal-to-noise ratio results of the proposed method with various numbers of input images.

Image #Input2 images4 images6 images8 images
127.6529.6131.432.1932.6
227.6429.5231.3732.1732.6
327.6231.3432.1932.65
427.5931.2732.1732.64
527.5932.1232.62
627.5732.0132.55
727.5432.48
827.5232.38
Average27.5929.5731.3532.1432.57

We also perform an experiment on two sets of real infrared images as shown in Fig. 4. We have collected the two sets of images using a 640×512 InSb focal plane array camera operating in the 3 to 5 μm range. One set of images in Fig. 4(a) shows drastic intensity changes among them due to atmospheric effects. The other set of images in Fig. 4(b) has relatively large motion. Objective results for the proposed method are provided in Table 2. We employ a roughness metric8,9,15 which is defined by

(18)

ρ=h1*I1+h2*I1I1,
where h1 and h2 are horizontal and vertical difference filters, respectively, I represents the image under test, 1 is the L1 norm, and * denotes discrete convolution. The roughness metric ρ measures the amount of high-frequency energy due to the RNU. As pointed out in Ref. 9, the metric cannot distinguish between true high-frequency energy and that from the RNU. However, the metric can be a useful indicator of the RNU to some degree when taken along with subjective evaluation.9 The correction results of the proposed method are depicted in Figs. 5 and 6. Similar to the simulated nonuniformity case, our method consistently suppresses the RNU as shown in Figs. 5(b) and 6(b). We also present the difference images between the input and the corrected in Fig. 7 to visualize the RNU corrected by the proposed method.

Fig. 4

Real infrared images.

OE_53_5_053105_f004.png

Table 2

Roughness results for real images.

Image #Fig. 4(a)Fig. 4(b)
InputProposedInputProposed
10.22110.07170.21830.0626
20.12900.07980.39810.0618
30.46180.06480.43380.0465
40.73890.0484
50.39750.0685

Fig. 5

NUC results on images shown in Fig. 4(a). (a) Whole images. (b) Close-ups of the input (top) and the corrected (bottom).

OE_53_5_053105_f005.png

Fig. 6

NUC results on images shown in Fig. 4(b). (a) Whole images. (b) Close-ups of the input (top) and the corrected (bottom).

OE_53_5_053105_f006.png

Fig. 7

Difference images between the input and the corrected for (a) the fifth image in Fig. 4(a) and the third image in Fig. 4(b).

OE_53_5_053105_f007.png

We have implemented the proposed method using C language. The simulation is performed on a PC with an Intel i7 3.40-GHz CPU and 4-GB memory. Our optimization procedure takes 5.6 and 3.8 s for the set of images in Figs. 4(a) and 4(b), respectively.

4.

Conclusion

We presented a regularization approach to SBNUC with several image frames. Our work formulated the SBNUC process as the energy minization problem that incorporates the slowly varying nature of the detector responsivity and the smoothness constraint for the scene irradiance. In the proposed method, no assumption was made about the motion among input images except that the motion is static. Therefore, the proposed method can be used in applications where only several image frames are available and large displacements exist among the given images. Simulation results on both synthetic and real infrared images demonstrated that the proposed method can reduce the RNU. In future works, we plan to apply more advanced numerical techniques to reduce the computational complexity of the proposed method.

References

1. D. L. PerryE. L. Dereniak, “Linear theory of nonuniformity correction in infrared staring sensors,” Opt. Eng. 32(8), 1854–1859 (1993).OPEGAR0091-3286 http://dx.doi.org/10.1117/12.145601 Google Scholar

2. W. GrossT. HierlM. Schulz, “Correctability and long-term stability of infrared focal plane arrays,” Opt. Eng. 38(5), 862–869 (1999).OPEGAR0091-3286 http://dx.doi.org/10.1117/1.602055 Google Scholar

3. L. Shkedyet al., “Megapixel digital InSb detector for midwave infrared imaging,” Opt. Eng. 50(6), 061008 (2011).OPEGAR0091-3286 http://dx.doi.org/10.1117/1.3572163 Google Scholar

4. C. ZhangW. Zhao, “Scene-based nonuniformity correction using local constant statistics,” JOSA A 25(6), 1444–1453 (2008).JOAOD61084-7529 http://dx.doi.org/10.1364/JOSAA.25.001444 Google Scholar

5. J. G. HarrisY.-M. Chiang, “Nonuniformity correction of infrared image sequences using the constant-statistics constraint,” IEEE Trans. Image Process. 8(8), 1148–1151 (1999).IIPRE41057-7149 http://dx.doi.org/10.1109/83.777098 Google Scholar

6. C. Zuoet al., “Scene-based nonuniformity correction method using multiscale constant statistics,” Opt. Eng. 50(8), 087006 (2011).OPEGAR0091-3286 http://dx.doi.org/10.1117/1.3610978 Google Scholar

7. D. A. Scribneret al., “Adaptive nonuniformity correction for ir focal-plane arrays using neural networks,” Proc. SPIE 1541, 100–109 (1991).PSISDG0277-786X http://dx.doi.org/10.1117/12.49324 Google Scholar

8. E. VeraS. Torres, “Fast adaptive nonuniformity correction for infrared focal-plane array detectors,” EURASIP J. Appl. Signal Process. 2005, 1994–2004 (2005).1110-8657 http://dx.doi.org/10.1155/ASP.2005.1994 Google Scholar

9. R. C. Hardieet al., “Scene-based nonuniformity correction with reduced ghosting using a gated lms algorithm,” Opt. Express 17(17), 14918–14933 (2009).OPEXFF1094-4087 http://dx.doi.org/10.1364/OE.17.014918 Google Scholar

10. A. RossiM. DianiG. Corsini, “Temporal statistics de-ghosting for adaptive non-uniformity correction in infrared focal plane arrays,” Electron. Lett. 46(5), 348–349 (2010).ELLEAK0013-5194 http://dx.doi.org/10.1049/el.2010.3559 Google Scholar

11. L. Bottou, “Large-scale machine learning with stochastic gradient descent,” in Proc. COMPSTAT’2010, pp. 177–186, Springer (2010). Google Scholar

12. C. Zuoet al., “Scene-based nonuniformity correction algorithm based on interframe registration,” JOSA A 28(6), 1164–1176 (2011).JOAOD61084-7529 http://dx.doi.org/10.1364/JOSAA.28.001164 Google Scholar

13. C. Zuoet al., “Improved interframe registration based nonuniformity correction for focal plane arrays,” Infrared Phys. Technol. 55(4), 263–269 (2012).IPTEEY1350-4495 http://dx.doi.org/10.1016/j.infrared.2012.04.002 Google Scholar

14. E. VeraP. MezaS. Torres, “Total variation approach for adaptive nonuniformity correction in focal-plane arrays,” Opt. Lett. 36(2), 172–174 (2011).OPLEDP0146-9592 http://dx.doi.org/10.1364/OL.36.000172 Google Scholar

15. C. Zuoet al., “A two-frame approach for scene-based nonuniformity correction in array sensors,” Infrared Phys. Technol. 60, 190–196 (2013).IPTEEY1350-4495 http://dx.doi.org/10.1016/j.infrared.2013.05.001 Google Scholar

16. C. Zuoet al., “Scene based nonuniformity correction based on block ergodicity for infrared focal plane arrays,” Optik 123(9), 833–840 (2012).OTIKAJ0030-4026 http://dx.doi.org/10.1016/j.ijleo.2011.06.050 Google Scholar

17. M. Maoet al., “Based on airborne multi-array butting for IRFPA staring imagery,” Proc. SPIE 7658, 765858 (2010). http://dx.doi.org/10.1117/12.865976 Google Scholar

18. C. R. del BlancoF. JaureguizarN. García, “Robust tracking in aerial imagery based on an ego-motion bayesian model,” EURASIP J. Adv. Signal Process. 2010, 1–18 (2010).1110-8657 http://dx.doi.org/10.1155/2010/837405 Google Scholar

19. E. GurevichA. Fein, “Maintaining uniformity of ir focal plane arrays by updating offset correction coefficients,” Proc. SPIE 4820, 809–820 (2003). http://dx.doi.org/10.1117/12.453552 Google Scholar

20. O. Nesheret al., “Digital cooled insb detector for ir detection,” Proc. SPIE 5074, 120–129 (2003).PSISDG0277-786X http://dx.doi.org/10.1117/12.498154 Google Scholar

21. Z. Farbmanet al., “Edge-preserving decompositions for multi-scale tone and detail manipulation,” ACM Trans. Graph. 27(3), 1–67 (2008).ATGRDF0730-0301 http://dx.doi.org/10.1145/1360612 Google Scholar

22. B. P. Flanneryet al., Numerical Recipes in C, Press Syndicate of the University of Cambridge, New York (1992). Google Scholar

23. G. Gershonet al., “3 Mega-pixel InSb detector with 10 μm pitch,” Proc. SPIE 8704, 870438 (2013).PSISDG0277-786X http://dx.doi.org/10.1117/12.2015583 Google Scholar

24. P. J. BexS. G. SolomonS. C. Dakin, “Contrast sensitivity in natural scenes depends on edge as well as spatial frequency structure,” J. Vision 9(10), 11–19 (2009).1534-7362 http://dx.doi.org/10.1167/9.10.1 Google Scholar

Biography

Jun-Hyung Kim received his BS and PhD degrees in electronic engineering from Korea University in 2006 and 2012, respectively. He has worked for the Agency for Defense Development since 2012. His current research interests are in the area of image processing and infrared imaging system. He is a member of SPIE.

Jieun Kim received her BS degree in electrical engineering from Busan National University in 2002, and her MS degree in electrical engineering from KAIST in 2004. She has worked for the Agency for Defense Development since 2004. Her current interests include digital image processing, target detection, and tracking.

Sohyun Kim is currently a research member at the Agency for Defense Development in Korea, and has over 10 years of experience in developing electro-optic systems. Her experience includes target detection algorithm design and real-time implementation for video tracker for infrared images. She holds a BS in physics from Sogang University and an MS in information and communications from Gwangju Institute of Science and Technology.

Joohyoung Lee received his BS and MS degress in electronics engineering from Dankuk University, Republic of Korea, in 1990 and 1992, respectively. Since 1992, he has been a principal researcher in the Electro-Optics Laboratory at the Agency for Defense Development. His research interests include analog and digital signal processing for IRST, low-noise electronics, IRST system test and evaluation for small target detection.

Boohwan Lee received his BS, MS, and PhD degrees in electrical engineering and computer science from Kyungpook National University, Daegu, Republic of Korea in 1991, 1993, and 2006, respectively. He has worked as a principal researcher for the Agency for Defense Development since 1993. His current interests include digital image processing, target detection, and tracking.

© The Authors. Published by SPIE under a Creative Commons Attribution 3.0 Unported License. Distribution or reproduction of this work in whole or in part requires full attribution of the original publication, including its DOI.
Jun-Hyung Kim, Jun-Hyung Kim, Jieun Kim, Jieun Kim, Sohyun Kim, Sohyun Kim, Joohyoung Lee, Joohyoung Lee, Boohwan Lee, Boohwan Lee, } "Regularization approach to scene-based nonuniformity correction," Optical Engineering 53(5), 053105 (7 May 2014). https://doi.org/10.1117/1.OE.53.5.053105 . Submission:
JOURNAL ARTICLE
7 PAGES


SHARE
Back to Top