Open Access
Add to BibSonomy
Abstract
In polarimetric imaging, degree and angle of linear polarization (DoLP and AoLP, respectively) are computed from ratios of Stokes parameters. In snapshot imagers, DoLP and AoLP are degraded by inherent mismatches between the spatial bandwidth of the S0, S1, and S2 parameters reconstructed by demosaicking from microgrid polarizer array (MPA)-sampled data. To overcome this, we rigorously show that log-MPA-sampled data approximately decouples DoLP and AoLP from the intensity component (S0) in the spatial Fourier domain. Based on this analysis, we propose an alternative demosaicking strategy aimed at estimating DoLP and AoLP directly from MPA-sampled data. Our method bypasses Stokes parameter estimation, alleviating the spatial bandwidth mismatch problems altogether and reducing computational complexity. We experimentally verify the superior DoLP and AoLP reconstructions of the proposed log-MPA demosaicking compared to the conventional Stokes parameter demosaicking approach in simulation. We simulated the conventional 2 × 2 MPA patterns as well as the more recently introduced 2 × 4 MPA patterns, and report quantitative results (mean squared error, structural similarity index, and polarization angular error) using five demosaicking approaches drawn from the literature. We also provide a closed-form error analysis on the log-MPA-sampled data to demonstrate that the approximation error is negligible for real practical applications.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
In remote sensing, both temporal and spatial modulation are common methods for inferring the polarization state of light. In snapshot imaging polarimetry, microgrid polarizer array (MPA) or division of focal plane (DoFP) polarimetry devices use repeating patterns of wiregrid polarizers bonded to a focal plane array [1–4]. One of the most commonly used microgrid patterns is a $2\times 2$ repeating pattern of pixel-sized polarization analyzers developed in 1994 and has been standard since its inception [5,6]. More recent $2\times 4$ microgrid arrangement boasts a higher reconstructed spatial resolution when compared to the standard $2\times 2$ pattern [7]. MPA examples are shown in Fig. 1.
Fig. 1. Micro-Polarizer Arrays (MPA) interrogated in this paper. The two pictured MPAs included are labeled as (a) 2x2 MPA (b) 2x4 MPA.
The ability to capture Stokes parameters in a single snapshot requires a post-capture interpolation step known as “demosaicking” to reconstruct the polarization states from MPA sampled data [8–15]. In applications that make use of degree and angle of linear polarization (DoLP and AoLP, respectively), the recovery of linear Stokes vector $\boldsymbol{S}=[S_0,S_1,S_2,S_3]^T\in \mathbb {R}^4$ [16] is only an intermediary step. Yet, existing MPA demosaicking methods are optimized for Stokes parameter recovery. Hence, the quality of the final DoLP/AoLP values computed from demosaicked Stokes vector are often suboptimal, stemming from the mismatch in power and spatial bandwidth supports of $S_0\in \mathbb {R}$, $S_1\in \mathbb {R}$, and $S_2\in \mathbb {R}$ parameters reconstructed by the MPA demosaicking algorithms. Rigorous details are provided in Section 2.3.
In this work, we propose a novel technique to re-purpose existing MPA demosaicking methods (intended for Stokes parameter reconstruction) to estimate DoLP and AoLP directly from MPA-sampled data. This work relies on a rigorous analysis of log-MPA-sampled sensor data showing that the logarithm function approximately decouples DoLP and AoLP from the intensity component, $S_0$, and that this approximation is valid across the span of DoLP values found in nature. We demonstrate that the demosaicking of log-MPA-sampled sensor data leads to direct DoLP/AoLP recovery, bypassing Stokes parameter estimation altogether. The proposed technique improves the accuracy of the estimated DoLP/AoLP values while also reducing computational complexity.
We experimentally verify the superior DoLP and AoLP reconstruction quality of the proposed log-MPA-based demosaicking compared to the conventional Stokes parameter demosaicking approach. We simulated the conventional $2\times 2$ MPA pattern as well as the more recently introduced $2\times 4$ MPA pattern. We report quantitative (peak signal-to-noise ratio, structural similarity index, Spearman rank correlation, and polarization angular error) and qualitative results to demonstrate the benefits to the novel log-MPA demosaicking method. In addition to empirical experiments, we derive an analytical form of the approximation error which is shown to play a negligible role in the real practical imaging scenarios.
2. Background
2.1 Representations of polarized light
Degree of linear polarization (DoLP) and angle of linear polarization (AoLP) are common terms for describing the linearly polarized component of incoherent light. In this way, light is decomposed into intensity, $I\in \mathbb {R^+}$; DoLP, $Y\in [0,1]$; and AoLP, $\theta \in \mathbb {R}/\pi$. Furthermore, the Stokes parameters are another canonical representation of polarization state for incoherent light. The complete set of parameters are represented in vector form $\boldsymbol{S}=[S_0,S_1,S_2,S_3]^T\in \mathbb {R}^4$, but the polarization analyzers in the MPAs are insensitive to circular polarization component $S_3$. Although circular polarization has been shown to exist in biological samples when analyzed with quarter waveplate polarization state analyzers [17,18] and therefore cannot be fully dismissed, $S_3=0$ is approximately true in many passive remote sensing applications [5]. Thus $S_3\in \mathbb {R}$ is suppressed for the remainder of this work. The relationships between the Stokes parameters, DoLP, and AoLP are given by:
Intuitively, the $S_0=I\in \mathbb {R^+}$ component is the intensity and $S_1,S_2\in \mathbb {R}$ are the linear polarization components.
Consider an intensity measurement made through a polarization analyzer. The measured optical intensity, $I_{\psi }\in \mathbb {R^+}$, through an analyzer oriented at an angle $\psi \in \mathbb {R}/\pi$ is an inner product between the analyzer vector, $\boldsymbol{A}_{\psi }\in \mathbb {R}^{3}$, and the Stokes vector:
Finally, DoLP and AoLP can be recovered from the Stokes vector in Eq. (3):
2.2 Microgrid polarizer array
MPA can be modelled as a spatially varying analyzer, $\boldsymbol{A}(\boldsymbol{n})=\boldsymbol{A}_{\psi (\boldsymbol{n})}$, where $\psi (\boldsymbol{n})$ is the orientation of the analyzer at pixel location $\boldsymbol{n} = [n_1,n_2]^T\in \mathbb {Z}^2$ ($n_1$ and $n_2$ are horizontal and vertical axes of the image, respectively). For example $2\times 2$ MPA-sampled sensor data $X(\boldsymbol{n})$ can be written as follows [7]:
It follows from Fourier analysis of Eq. (5) and Eq. (6) in Fig. 3 that MPA sampling results in a modulation of $S_1$ and $S_2$ components to carrier frequencies $[\pi,0]$ and $[0,\pi ]$ for $2\times 2$ MPA, and $[\pm \pi /2,\pi ]$ for $2\times 4$ MPA.
Broadly speaking, there are three categories of demosaicking approaches to MPA sampled data, as shown in Fig. 2. The first approach is to recover a full-resolution analyzer images $\{I_{0},I_{\pi /4},I_{\pi /2},I_{3\pi /4}\}$ from the MPA sampled data [9–11]. The recovered analyzer images can be used to compute Stokes parameter using Eq. (3) and DoLP/AoLP using Eq. (4). The second approach is to interpret demosaicking task as a demodulation of the Stoke parameters $\{S_0,S_1,S_2\}$ directly [8]. Demodulation is achieved by applying a lowpass filter to MPA sampled data multiplied by the carrier frequency, depicted in Fig. 3. The estimated Stokes parameter are subsequently used in Eq. (4) to compute DoLP/AoLP. We propose a third approach where the demosaicking estimates DoLP/AoLP directly.
Fig. 2. System diagrams describing three workflows for recovering DoLP/AoLP from MPA-sampled-data. Demosaicking with filters designed to match the spatial bandwidths of the analyzer images or Stokes parameters may not yield the best DoLP/AoLP reconstruction. The proposed log-MPA-sampled demosaicking allows filters to match the spatial bandwidth of DoLP/AoLP.
Fig. 3. Fourier domain representations of Stokes parameters $\{S_0,S_1,S_2\}$ in the MPA-sampled sensor data, computed from “bubblewrap” image in the dataset [19]. (a) $2\times 2$ MPA (b) $2\times 4$ MPA.
2.3 Problems with bandwidth mismatch in demosaicking
In conventional snapshot polarimetry imaging, Stokes parameters are first recovered by demosaicking. The demosaicked Stokes parameters are subsequently used to compute DoLP and AoLP by the relation in Eq. (4). As stated in the introduction, however, the reconstruction quality of DoLP and AoLP are suboptimal. The deterioration stems from the fact that the power spectra of the $S_1$ and $S_2$ components are much smaller than that of the baseband instensity signal $S_0$—the signal power is only 10% of the baseband in typical long-wave infrared imaging scenarios, for example. Hence the demosaicking-based Stokes vector reconstruction is an asymmetrical signal reconstruction task because the estimated intensity image signal $S_0$ typically supports a higher spatial bandwidth with greater signal magnitude than that of the polarized images $S_1$ and $S_2$.
This inherent mismatch between the spatial bandwidth supports of demosaicked $S_0$, $S_1$, and $S_2$ parameters degrades the DoLP image $Y({\boldsymbol{n}})$ computed from the Stokes parameter ratios. To understand why this is the case, consider rewriting Eq. (4) as:
Table 1. Mean squared error (MSE) of reconstructed Stokes parameter $S_0$ and DoLP $Y$, computed with lowpass filter applied to $S_0$ demosaicked by the method in [8]. The average error for all images in the dataset [19] is reported. The spatial bandwidth of $S_0$ is varied while the bandwidths of $S_1$ and $S_2$ are kept constant at $0.4\pi$ radius. Limiting the spatial bandwidth of $S_0$ to $0.3\pi$ radius improves the DoLP reconstruction quality, even though the Stokes parameter reconstruction is worse.
3. Proposed DoLP/AoLP recovery from Log-MPA data
3.1 DoLP/AoLP Analysis of MPA Sampling
Consider rewriting the analyzer in Eq. (2) in terms of intensity $I$, DoLP $Y$, and AoLP $\theta$, as follows:
The same process used in Eq. (9) gives rise to the DoLP/AoLP analysis of $2\times 4$ MPA sampling:
3.2 Analysis of Log-MPA sampling
One main feature of Eq. (9) and Eq. (10) is that the intensity image $I(\boldsymbol{n})$ is factored out. This identity is particularly well-matched for the following logarithm identity and approximation:
This well-known approximation of $\ln (1+\frac {v}{u})$ stems from the first order Taylor series expansion. In Section 4, we provide a rigorous full Taylor series expansion analysis to assess the practical impact of approximation error on the demosaicking results.
Suppose we let $u=I(\boldsymbol{n})/2$, and interpret $v/u$ as the modulated DoLP/AoLP terms in Eq. (9) and Eq. (10). When $Y(\boldsymbol{n})\ll 1$, the logarithm approximation in Eq. (11) applies:
The above log-MPA analysis is significant because it linearizes the DoLP/AoLP values. Specifically, the logarithm approximately decouples the baseband log-intensity $\ln (I(\boldsymbol{n})/2)$ from the DoLP/AoLP terms $Y(\boldsymbol{n})e^{\pm j2\theta (\boldsymbol{n})}$. Furthermore, DoLP/AoLP terms are modulated by $(-1)^{n_1}$ and $(-1)^{n_2}$ in the case of $2\times 2$ MPA, and by $\exp (\pm j(\frac {\pi }{2}n_1+\pi n_2))$ in the case of $2\times 4$ MPA. The log-intensity image remains at the baseband (i.e. not modulated).
Figure 4 depicts the Fourier domain representations of $\ln (X(\boldsymbol{n}))$ for both the $2 \times 2$ and $2 \times 4$ MPAs. Indeed, the concentration of energy around the modulation carrier frequencies $[0,\pi ]$ and $[\pi,0]$ in Fig. 4(a) and $[\pm \pi /2,\pi ]$ in Fig. 4(b) correspond to the DoLP/AoLP values of interest. They are well-separated from the baseband log-intensity image signal $\ln (I(\boldsymbol{n})/2)$.
Fig. 4. Fourier domain representations of $Y(\boldsymbol{n})$ and $\theta (\boldsymbol{n})$ in the log-MPA-sampled sensor data, computed from “bubblewrap” image in the dataset [19]. (a) 2$\times$2 MPA (b) 2$\times$4 MPA.
3.3 Proposed: demosaicking of Log-MPA sampled data
The log-MPA analysis in Section 3.2 yielded a modulation in the Fourier domain as shown in Fig. 4. This is similar to the Stokes analysis of MPA sampling in Section 2.2 and Fig. 3, save for the fact that the modulated signals are combinations of $Y(\boldsymbol{n})e^{\pm j2\theta (\boldsymbol{n})}$ instead of $S_1$ and $S_2$. These similarities raise the possibility that an existing “Stokes parameter demosaicking” technique can be repurposed as direct “DoLP/AoLP demosaicking” with only a small modification.
Denote the Stokes parameter demosaicking operator as follows:
where $\{\hat {S}_0,\hat {S}_1,\hat {S}_2\}$ are the estimates of $\{S_0,S_1,S_2\}$. Without loss of generality, the above may also refer to Stokes parameters computed by Eq. (3) from demosaicked analyzer images $\{\hat {I}_{0},\hat {I}_{\pi /4},\hat {I}_{\pi /2},\hat {I}_{3\pi /4}\}$. Consider applying this demosaicking operator on log-MPA image $\ln (X(\boldsymbol{n}))$ instead of MPA image $X(\boldsymbol{n})$:Recall the relationship between Stokes and DoLP/AoLP in Eq. (1). Substituting the logarithmic approximation in Eq. (11) where we set the baseline signal to $u=I(\boldsymbol{n})/2$, the demosaicking would be applied to recover the modulated signal in $\frac {v}{u}=\frac {X(\boldsymbol{n})}{I(\boldsymbol{n})/2}$. That is, applying demosaicking to log-MPA would yield
Thus, the following post-demosaicking processing would recover the DoLP $Y(\boldsymbol{n})$ and AoLP $\theta (\boldsymbol{n})$ as magnitude and angles of the vector $[\hat {T}_1,\hat {T}_2]/2$:
Taking Eq. (16) and Eq. (17) together, we have a plug-and-play approach to repurposing an existing MPA demosaicking algorithm for DoLP and AoLP recovery. This leads to a superior DoLP/AoLP reconstruction since DoLP/AoLP terms in Eq. (12) and Eq. (13) (which is recovered by demosaicking in Eq. (16)) were decoupled from $S_0$ term by applying logarithmic function to MPA (instead of computing ratios of Stokes parameters with mismatched spatial bandwidths). Our experiments in Section 5 confirm this claim. We note however that $\hat {T}_0$ is an unnecessary quantity if the end goal is to yield DoLP and/or AoLP. If $I$ needs to be recovered also, it takes an additional computational step to reconstruct $\hat {T}_0$ when compared to the conventional demosaicking in Eq. (14) which estimates $I=S_0$ directly. Consequently, it is most computationally efficient to estimate $S_0$ in the conventional way with Eq. (14) and then DoLP/AoLP with Eq. (17).
4. Error analysis
The basis of the log-MPA analysis presented in the previous section relies on a first order Taylor series approximation. In this section, we provide an analysis of its approximation error. The Taylor series expansion of the $\ln (1+x)$ is defined as
Recalling the $2\times 2$ log-MPA analysis in Eq. (12), suppose we substitute
Fig. 5. Fourier domain representations of approximation stemming from Taylor series expansion of logarithm function, computed from “bubblewrap” image in the dataset [19]. (a) 2$\times$2 MPA (b) 2$\times$4 MPA.
Let us repeat the same analysis on the $2\times 4$ MPA pattern. Substituting
As shown by Fig. 5(b), the odd error term in Eq. (24) modulates to $[\pm \frac {\pi }{2},\pi ]$; and the same analysis leads to the even error term modulated to $[0,0]$ or $[\pi,0]$.
Recall that the demosaicking task can be interpreted as the demodulation of $S_1$ and $S_2$ signals near carrier frequencies $[0,\pi ]$ and $[\pi,0]$ for $2\times 2$ MPA pattern (or $[\pm \frac {\pi }{2},\pi ]$ for $2\times 4$ MPA pattern). Thus the even error terms modulated to $[0,0]$ and $[\pi,\pi ]$ in $2\times 2$ MPA pattern (or to $[0,0]$ or $[\pi,0]$ in $2\times 4$ pattern) are filtered out and suppressed by default—compare Fig. 5 with Fig. 3 and Fig. 4. Thus only the odd error term in Eq. (20) contributes to the overall degradation of the final demosaicking result.
One can solve for the odd error term in Eq. (20) explicitly as
using the relationRegarding Eq. (26) as the contributing error term, the corresponding DoLP to approximation error ratio in decibels can be computed by:
Figure 6 shows this error approximation ratio as a function of DoLP. For long wave infrared (LWIR), the typical DoLP signal is between the range of 0.1% to 20% [5,20,21]—at this range, the approximation performs very well, with a minimum approximation error ratio of 37.29 dB. For visible, near infrared (NIR), and short-wave infrared (SWIR) wavelength regions, the DoLP signal may be as high as 60% [5], corresponding to approximation error ratio of 16.18 dB.
5. Experimental results
In this section, we experimentally compare the conventional linear demosaicking approach $X(\boldsymbol{n})\rightarrow \operatorname {\Gamma }(X(\boldsymbol{n}))\rightarrow \operatorname {Stokes} \rightarrow Y(\boldsymbol{n})$ to the proposed log-MPA-sampled demosaicking method $\ln (X(\boldsymbol{n}))\rightarrow \operatorname {\Gamma }(\ln (X(\boldsymbol{n})))\rightarrow Y(\boldsymbol{n})$. For this study, we simulate MPA from the dataset in [19] using the average of the color channels and report objective performance metrics averaged over the entire dataset in Tables 2–6. We then show that existing demosaicking approaches, such as demodulation [8], NLPN [9], bicubic and bicubic spline [10], Adaptive [11], and Newton-Poly [12] interpolation algorithms, can be applied to log-MPA data to directly estimate DoLP/AoLP without bandwidth mismatch errors. Note that the $2\times 4$ MPA pattern has limited number of demosaicking algorithms compared to the $2\times 2$ MPA pattern due to the fact that it was introduced more recently. We excluded CNN-based demosaicking algorithms such as [13–15] from our experiments because the dynamic ranges of the log-MPA samples are significantly different from linear-MPAs. Getting these methods to operate properly on log-MPA requires retraining.
Table 2. Peak signal-to-noise ratio (PSNR) of DoLP reconstructed from several demosaicking techniques applied to linear-MPA-sampled as well as (proposed) log-MPA-sampled sensor data.
Table 3. Structural similarity index (SSIM) of DoLP reconstructed from several demosaicking techniques applied to linear-MPA-sampled as well as (proposed) log-MPA-sampled sensor data.
Table 4. Spearman rank correlation of DoLP reconstructed from several demosaicking techniques applied to linear-MPA-sampled as well as (proposed) log-MPA-sampled sensor data.
Table 5. Spectral angle mapper of AoLP reconstructed from several demosaicking techniques applied to linear-MPA-sampled as well as (proposed) log-MPA-sampled sensor data.
Table 6. Mean squared DoLP-orthogonal error in Eq. (30) of AoLP reconstructed from several demosaicking techniques applied to linear-MPA-sampled as well as (proposed) log-MPA-sampled sensor data.
We apply demosaicking to log-MPA, and compute DoLP and AoLP using Eq. (17). We compare this to the conventional method of applying demosaicking to MPA, and subsequently computing DoLP and AoLP from the intermediary recovered Stokes parameters $\{\hat {S}_0,\hat {S}_1,\hat {S}_2\}$ using Eq. (4). We assess the DoLP reconstruction qualities based on peak signal-to-noise ratio (PSNR) (Table 2), structural similarity index measure (SSIM) (Table 3), and Spearman rank correlation. Spearman rank correlation of DoLP (Table 4) is used to measure the consistency between the ground truth DoLP and reconstructed DoLP when the DoLP images are thresholded (which is often how DoLP images are used in practice).
For AoLP assessment, we show AoLP error as spectral angle mapper (SAM) (Table 5) defined as:
Recalling that DoLP and AoLP form a polar coordinate pair, DoLP errors in Table 2–4 are errors in radial directions. AoLP error contributes to the component of the estimation error that is orthogonal to the radial direction. One can compute the magnitude of DoLP-orthogonal error component as
Notice that AoLP error is weighted by DoLP, meaning that a large AoLP error value in a low DoLP region does not contribute much to the overall error. This weighting ensures that AoLP estimates in unpolarized regions of the image (where noise dominates and AoLP has no physical meaning) do not skew the results. We caution that the quantity in Eq. (30) is vector magnitude, and not a vectored value. The fact that DoLP $Y$ (magnitude of the radial component, also not a vector) appears in Eq. (30) does not mean that the two vectors are not orthogonal. We report the mean squared DoLP-orthogonal error in Table 6.
With few exceptions, DoLP and AoLP reconstructions from the proposed log-MPA demosaicking is superior to the conventional linear-MPA demosaicking approach. This result suggests that the approximation error shown in Fig. 6 is not a dominant source of distortion. In each assessment category, demodulation-based demosaicking in [8] had the best performance. Demodulation, bicubic, and bicubic spline demosaicking methods were consistently improved by the proposed log-MPA approach. The demodulation approach employs a filter size of $0.4\pi$, which has been validated in Section 2.3 to have the lowest MSE for $S_0$ recovery. As a side note, the objective score improvements by log-MPA demosaicking in Tables 2–6 may seem modest in comparison to non-uniformity correction [22–25]. The latter is by-and-large a pixel-wise operation that is not designed to improve the resolution of the recovered high frequency components, however, and thus our work in demosaicking is complimentary to non-uniformity correction.
Example DoLP images reconstructed from linear and log demosaicking are shown in Fig. 7, using the demodulation demosaicking technique in [8] that performed best in our quantitative evaluation. None of the reconstructions are perfect, as evidenced by aliasing artifacts. However, the log-MPA approach suffers from less artifacts compared to the linear-MPA counterparts. This is particularly true for regions of the image where there is a sharp transition between high and low DoLP values, where the linear-MPA approach seems to suffer from ringing artifacts extending into the low DoLP regions. For example, parallel horizontal lines in the middle column image that appear merged or jagged in linear-MPA demosaicking results are resolved better with log-MPA demosaicking. The log-MPA demosaicking seems to be more robust to ringing artifacts at strong edge regions, also.
Fig. 7. Example DoLP images using demodulation demosaicking [8] and simulated MPAs from dataset in [19].
Example AoLP images in Fig. 8 are more difficult to interpret. Again, none of the demosaicking methods yield perfect results. In some regions, log-MPA-based AoLP reconstructions appear to be more consistent with the ground truth than the linear-MPA reconstructions,such as the ringing artifacts in the right column images, and the erroneous thin striped AoLP values in the left column images. Comparing to Fig. 7, however, the regions where linear-MPA does better in AoLP reconstruction than log-MPA seem to also coincide with low DoLP values, such as the ringing artifacts in the background next to the bottle image. In other words, AoLP fidelity of log-MPA reconstruction remains high where the polarization signatures are strong, a fact that is also supported by the quantitative assessment in Table 6 (AoLP error weighted by DoLP). We contrast this to the linear-MPA reconstruction, where the AoLP error coincides with DoLP ringing error in Fig. 7. This is potentially hazardous since falsely high DoLP values exhibit false AoLP features as well.
Fig. 8. Example AoLP images using demodulation demosaicking [8] and simulated MPAs from dataset in [19]. Angle $\hat {\theta }(\boldsymbol{n})$ shown as hue angle.
6. Conclusion
In this paper, we rigorously prove that DoLP/AoLP parameters are approximately decoupled in log-MPA-sampled sensor data. This approximation is then used to propose a novel framework to estimate DoLP and AoLP from demosaicking of log-MPA-sampled sensor data. Specifically, conventional MPA demosaicking algorithms designed for estimating Stokes parameters may be repurposed to yield DoLP and AoLP by operating on log-MPA-sampled sensor data instead of the MPA-sampled sensor data. By bypassing Stokes parameter estimation, we arrive at superior DoLP/AoLP reconstructions that avoid the pitfalls of the spatial frequency mismatches between the baseband $S_0$ and the linear polarization components $S_1$ and $S_2$. We also analytically derived the approximation error, and proved that it has negligible impact on the accuracy of the recovered DoLP/AoLP.
In experiments, we simulated 2$\times$2 and 2$\times$4 MPA sampling. We evaluated the reconstruction qualities of the DoLP/AoLP computed from Stokes parameters reconstructed by demosaicking, as well as the accuracies of DoLP/AoLP from the demosaicking of log-MPA-samlped sensor data. For most choices of demosaicking algorithms (especially the best performing ones), the log-MPA demosaicking outperformed MPA demosaicking. We conclude that the proposed framework yields a way to optimize demosaicking task for the DoLP/AoLP values rather than Stokes parameters.
Funding
Air Force Research Laboratory (FA8650-21-F-1125).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are available in [19].
References
1. A. G. Andreou and Z. K. Kalayjian, “Polarization imaging: principles and integrated polarimeters,” IEEE Sens. J. 2(6), 566–576 (2002). [CrossRef]
2. G. P. Nordin, J. T. Meier, P. C. Deguzman, and M. W. Jones, “Diffractive optical element for stokes vector measurement with a focal plane array,” Proc. SPIE 3754, 169–177 (1999). [CrossRef]
3. C. K. Harnett and H. G. Craighead, “Liquid-crystal micropolarizer array for polarization-difference imaging,” Appl. Opt. 41(7), 1291–1296 (2002). [CrossRef]
4. L. B. Wolff and A. G. Andreou, “Polarization camera sensors,” Image Vis. Comput. 13(6), 497–510 (1995). [CrossRef]
5. J. S. Tyo, D. L. Goldstein, D. B. Chenault, and J. A. Shaw, “Review of passive imaging polarimetry for remote sensing applications,” Appl. Opt. 45(22), 5453–5469 (2006). [CrossRef]
6. C. S. Chun, D. L. Fleming, and E. Torok, “Polarization-sensitive thermal imaging,” Proc. SPIE 2234, 275–286 (1994). [CrossRef]
7. D. A. LeMaster and K. Hirakawa, “Improved microgrid arrangement for integrated imaging polarimeters,” Opt. Lett. 39(7), 1811–1814 (2014). [CrossRef]
8. J. S. Tyo, C. F. LaCasse, and B. M. Ratliff, “Total elimination of sampling errors in polarization imagery obtained with integrated microgrid polarimeters,” Opt. Lett. 34(20), 3187–3189 (2009). [CrossRef]
9. B. M. Ratliff, C. F. LaCasse, and J. S. Tyo, “Interpolation strategies for reducing ifov artifacts in microgrid polarimeter imagery,” Opt. Express 17(11), 9112–9125 (2009). [CrossRef]
10. S. Gao and V. Gruev, “Bilinear and bicubic interpolation methods for division of focal plane polarimeters,” Opt. Express 19(27), 26161–26173 (2011). [CrossRef]
11. B. M. Ratliff, C. F. LaCasse, and J. S. Tyo, “Adaptive strategy for demosaicing microgrid polarimeter imagery,” in 2011 Aerospace Conference, (IEEE, 2011), pp. 1–9.
12. N. Li, Y. Zhao, Q. Pan, and S. G. Kong, “Demosaicking dofp images using newton’s polynomial interpolation and polarization difference model,” Opt. Express 27(2), 1376–1391 (2019). [CrossRef]
13. J. Zhang, J. Shao, H. Luo, X. Zhang, B. Hui, Z. Chang, and R. Liang, “Learning a convolutional demosaicing network for microgrid polarimeter imagery,” Opt. Lett. 43(18), 4534–4537 (2018). [CrossRef]
14. G. C. Sargent, B. M. Ratliff, and V. K. Asari, “Conditional generative adversarial network demosaicing strategy for division of focal plane polarimeters,” Opt. Express 28(25), 38419–38443 (2020). [CrossRef]
15. J. Song, R. Abay, J. S. Tyo, and A. S. Alenin, “Transcending conventional snapshot polarimeter performance via neuromorphically adaptive filters,” Opt. Express 29(12), 17758–17774 (2021). [CrossRef]
16. G. G. Stokes, “On the composition and resolution of streams of polarized light from different sources,” Transactions of the Camb. Philos. Soc. 9, 399 (1851).10.1017/CBO9780511702266.010
17. C. L. Patty, D. A. Luo, F. Snik, F. Ariese, W. J. Buma, I. L. Ten Kate, R. J. van Spanning, W. B. Sparks, T. A. Germer, G. Garab, and M. W. Kudenov, “Imaging linear and circular polarization features in leaves with complete mueller matrix polarimetry,” Biochimica et Biophys. Acta (BBA)-General Subj. 1862(6), 1350–1363 (2018). [CrossRef]
18. C. U. Keller, F. Snik, C. L. Patty, D. Klindžic, M. Krasteva, D. S. Doelman, T. Wijnen, V. Pallichadath, D. M. Stam, B.-O. Demory, J. G. Kühn, H. J. Hoeijmakers, A. Pommerol, and O. Poch, “Design of the life signature detection polarimeter lsdpol,” Proc. SPIE 11443, 684–693 (2020).
19. M. Morimatsu, Y. Monno, M. Tanaka, and M. Okutomi, “Monochrome and color polarization demosaicking using edge-aware residual interpolation,” in 2020 IEEE International Conference on Image Processing (ICIP), (IEEE, 2020), pp. 2571–2575.
20. M. Kupinski, R. Chipman, and E. Clarkson, “Relating the statistics of the angle of linear polarization to measurement uncertainty of the stokes vector,” Opt. Eng. 53(11), 113108 (2014). [CrossRef]
21. J. S. Tyo, B. M. Ratliff, and A. S. Alenin, “Adapting the hsv polarization-color mapping for regions with low irradiance and high polarization,” Opt. Lett. 41(20), 4759–4762 (2016). [CrossRef]
22. D. L. Bowers, J. K. Boger, D. Wellems, S. Ortega, M. P. Fetrow, J. E. Hubbs, W. T. Black, B. M. Ratliff, and J. S. Tyo, “Unpolarized calibration and nonuniformity correction for long-wave infrared microgrid imaging polarimeters,” Opt. Eng. 47(4), 046403 (2008). [CrossRef]
23. J. Zhang, H. Luo, B. Hui, and Z. Chang, “Non-uniformity correction for division of focal plane polarimeters with a calibration method,” Appl. Opt. 55(26), 7236–7240 (2016). [CrossRef]
24. H. Fei, F.-M. Li, W.-C. Chen, R. Zhang, and C.-S. Chen, “Calibration method for division of focal plane polarimeters,” Appl. Opt. 57(18), 4992–4996 (2018). [CrossRef]
25. J. K. Boger, J. S. Tyo, B. M. Ratliff, M. P. Fetrow, W. T. Black, and R. Kumar, “Modeling precision and accuracy of a lwir microgrid array imaging polarimeter,” Proc. SPIE 5888, 58880U (2005). [CrossRef]
