Robust multicomponent AM-FM image demodulation: stability at energy-operator singularities, quantified accuracy beyond narrowband assumptions, and principled multiscale filterbank fusion

The 1994 multidimensional energy-separation algorithm (DESA) of Maragos and Bovik demodulates images modeled as amplitude-modulation/frequency-modulation (AM-FM) signals. Its published analysis acknowledges three limitations: outlier instantaneous-frequency estimates wherever the discrete energy operator output is near zero (mitigated only by 2D median filtering), unquantified accuracy loss when the slow-modulation/narrowband assumptions fail on real images, and a per-axis frequency estimation ceiling of one quarter of the sampling frequency; the fourth open issue — heuristic filterbank design and cross-scale fusion — is documented by the direct lineage rather than by the 1994/1995 source itself. Three decades of follow-up work resolved the frequency-ceiling item in practice (arccos- and analytic-signal-based estimators), but the remaining core is open: a principled multicomponent AM-FM image demodulation framework that is provably stable at near-singular pixels without relying on heuristic median post-filtering, carries explicit error bounds quantified under specified violations of the standard approximation assumptions on real images, and rests on a filterbank design and fusion rule justified by an explicit optimality or error criterion. This core is repeatedly and explicitly deferred across the direct lineage (1996, 2000, 2010, 2012 review) and remains unmet by the most recent regularized (2021) and higher-order energy-operator (2022) methods.

Background

A locally narrowband image texture can be modeled as a two-dimensional amplitude-modulation/frequency-modulation (AM-FM) signal $f(m,n)=a(m,n)\cos\phi(m,n)$, where the instantaneous amplitude $a(m,n)$ captures local contrast and the instantaneous frequency (IF) vector $\nabla\phi(m,n)$ captures emergent spatial oscillation; real images are treated as sums of such components isolated by a multiscale bandpass filterbank (originally Gabor filters, i.e. bandpass filters localized jointly in space and spatial frequency). The Teager–Kaiser energy operator, in one dimension $\Psi_d[x](n)=x^2(n)-x(n-1)x(n+1)$, approximately tracks the energy product $a^2\omega^2$ of an AM-FM signal; Maragos, Kaiser, and Quatieri (IEEE Trans. Signal Processing, 1993) built the one-dimensional energy separation algorithm (ESA) on it. Maragos and Bovik (ICIP 1994; archival version in J. Opt. Soc. Am. A 12(9):1867, 1995) extended this to images with the two-dimensional discrete energy operator $\Phi_d[f](m,n)=2f^2(m,n)-f(m-1,n)f(m+1,n)-f(m,n-1)f(m,n+1)$ and the multidimensional discrete energy separation algorithm (DESA), which estimates each IF component through an $\arcsin$ of a ratio involving division by $\Phi_d[f]$. This construction has three defects acknowledged in the source itself, plus a fourth open issue documented by the follow-up lineage. First, wherever $\Phi_d[f]$ is zero or nearly zero (low-magnitude regions, carrier-phase cancellations), the ratios blow up or their $\arcsin$ arguments exceed one, producing outlier frequency estimates that the authors mitigated only by 2D median filtering. Second, the derivation assumes slowly varying amplitude and frequency and bandlimited modulating signals — the quasi-eigenfunction approximation (QEA) conditions; how accuracy degrades when real bandpass image components violate them (sharp contrasts, rapid frequency variation) was not quantified. Third, because the estimates use $\arcsin$, each IF component is estimable only in $(0,\pi/2)$ radians/sample, one quarter of the sampling frequency, although the signal model assumes discrete-space frequencies over the full $[0,\pi]$ range per axis. Fourth, full-image demodulation requires decomposing the image with a filterbank and fusing per-channel estimates; the 1994/1995 papers simply adopt the pre-designed Gabor filterbank of Bovik et al. and give no design or fusion criterion, and the need for a principled one was made explicit only by the follow-up lineage. Later work developed the surrounding machinery: Havlicek, Harding, and Bovik (IEEE Trans. Image Processing, 1996, 2000) introduced multicomponent AM-FM representations with Kalman-based component tracking (a standard recursive state-estimation filter following each component's amplitude and phase from pixel to pixel), dominant component analysis (selection of the filterbank channel with the strongest response at each pixel), and the QEA with approximation-error analysis; Sivley and Havlicek (ICIP 2006) studied perfect-reconstruction AM-FM models; Murray, Rodriguez, and Pattichis (IEEE Trans. Image Processing, 2010) designed flat-passband equiripple multiscale filterbanks and the variable-spacing local-linear-phase (VS-LLP) estimator, which replaces $\arcsin$ with an $\arccos$ of shifted samples of the unit-modulus analytic signal at an integer spacing chosen for numerical conditioning, thereby extending practical IF estimation beyond the $(0,\pi/2)$ range; the analytic signal of a real signal is the complex signal whose real part is the signal, obtained by adding the quadrature (Hilbert-transform-like) component. In parallel, regularized optimization approaches to multicomponent demodulation matured in one dimension (Venkitaraman and Seelamantula 2014; Hu, Peng, Guo, and Xu, Signal Processing, 2021). What no inspected work provides is the joint objective synthesized from the source's acknowledged constraints and the lineage's open design issues: demodulation of real multicomponent images that is stable at the near-singular pixels, accurate with quantified bounds when the QEA assumptions fail, of unrestricted frequency coverage, and equipped with a filterbank design and fusion rule justified by an explicit optimality criterion rather than heuristics.

Problem Statement

Construct a principled framework for multicomponent AM-FM demodulation of real digital images — that is, for estimating, at every pixel, the instantaneous amplitudes and frequency vectors of the AM-FM components into which the image is decomposed by a multiscale filterbank — that jointly meets the following four requirements, a synthesis, for the same full-image demodulation task, of the constraints and observations explicitly stated in the 1994/1995 source (requirements 1-3) and of the open design issues documented in its direct lineage (requirement 4). (1) Singular-region stability: the estimation must be well-defined and stable at pixels where the discrete energy operator output (or the equivalent denominator quantity of the chosen estimator) is zero or near zero, replacing heuristic post-hoc median filtering with a systematically justified mechanism, accompanied by an explicit bound — rather than a mere monotonicity or trend statement, and not a bound obtained by clamping the estimate to a fixed or trivial value at or near the singularities — on the outlier rate or error incurred in exactly those regions. (2) Quantified accuracy under assumption violations: instantaneous amplitude and frequency error bounds must be stated and verifiable in explicit functional form under explicitly specified violations of the standard slow-modulation, narrowband (quasi-eigenfunction-approximation) conditions — sharp contrasts and rapid frequency variation of the kind occurring in natural bandpass image components — not only in the asymptotic ideal regime, with the parameterized violation regime under which each bound holds stated explicitly rather than reporting only that the error grows. (3) Full frequency coverage: the framework must estimate each instantaneous-frequency component over the full discrete frequency range per axis — up to the per-axis Nyquist limit of $\pi$ radians/sample — rather than being intrinsically confined to $(0,\pi/2)$ radians/sample as the arcsin-based 2D DESA is, with estimation over the full per-axis range demonstrated rather than merely asserted. (4) Principled filterbank design and fusion: the analysis filterbank and the rule for fusing per-channel estimates across scales and components must be governed by an explicit optimality or error criterion that is shown to imply a component-separation or reconstruction-fidelity property — not a criterion merely named without one, such as naming 'minimize reconstruction mean-squared error' without any separation, selection, or fidelity property established — so that component separation, cross-scale selection, and reconstruction fidelity are justified rather than tuned ad hoc. Either construct such a joint framework, or determine and prove which of the four requirements blocks a joint solution.

References

1. P. Maragos and A. C. Bovik, 'Demodulation of images modeled by amplitude-frequency modulations using multidimensional energy separation', Proceedings of 1st International Conference on Image Processing (ICIP-94), 1994, pp. 421-425. DOI: 10.1109/ICIP.1994.413772, https://doi.org/10.1109/icip.1994.413772
2. P. Maragos and A. C. Bovik, 'Image demodulation using multidimensional energy separation', Journal of the Optical Society of America A 12(9), 1995, pp. 1867-1876. DOI: 10.1364/JOSAA.12.001867, https://doi.org/10.1364/JOSAA.12.001867 (archival journal version of the source)
3. P. Maragos, J. F. Kaiser, and T. F. Quatieri, 'On amplitude and frequency demodulation using energy operators', IEEE Transactions on Signal Processing 41(4), 1993, pp. 1532-1550. DOI: 10.1109/78.212729, https://doi.org/10.1109/78.212729
4. J. P. Havlicek, D. S. Harding, and A. C. Bovik, 'The multicomponent AM-FM image representation', IEEE Transactions on Image Processing 5(6), 1996, pp. 1094-1100. DOI: 10.1109/83.503927, https://doi.org/10.1109/83.503927
5. J. P. Havlicek, D. S. Harding, and A. C. Bovik, 'Multidimensional quasi-eigenfunction approximations and multicomponent AM-FM models', IEEE Transactions on Image Processing 9(2), 2000, pp. 227-242. DOI: 10.1109/83.821736, https://doi.org/10.1109/83.821736
6. R. A. Sivley and J. P. Havlicek, 'Perfect reconstruction AM-FM image models', 2006 International Conference on Image Processing, 2006, pp. 2125-2128. DOI: 10.1109/ICIP.2006.312828, https://doi.org/10.1109/icip.2006.312828
7. V. Murray, P. Rodriguez, and M. S. Pattichis, 'Multiscale AM-FM demodulation and image reconstruction methods with improved accuracy', IEEE Transactions on Image Processing 19(5), 2010, pp. 1138-1152. DOI: 10.1109/TIP.2010.2040446, https://doi.org/10.1109/tip.2010.2040446
8. V. Murray, M. S. Pattichis, E. S. Barriga, and P. Soliz, 'Recent multiscale AM-FM methods in emerging applications in medical imaging', EURASIP Journal on Advances in Signal Processing 2012:23, 2012. DOI: 10.1186/1687-6180-2012-23, https://doi.org/10.1186/1687-6180-2012-23
9. A. Venkitaraman and C. S. Seelamantula, 'Binaural signal processing motivated generalized analytic signal construction and AM-FM demodulation', IEEE/ACM Transactions on Audio, Speech, and Language Processing 22(6), 2014, pp. 1023-1036. DOI: 10.1109/TASLP.2014.2316376, https://doi.org/10.1109/taslp.2014.2316376
10. Y. Qin, 'Multicomponent AM-FM demodulation based on energy separation and adaptive filtering', Mechanical Systems and Signal Processing 38(2), 2013, pp. 440-459. DOI: 10.1016/j.ymssp.2013.01.023, https://doi.org/10.1016/j.ymssp.2013.01.023
11. X. Hu, S. Peng, B. Guo, and P. Xu, 'Accurate AM-FM signal demodulation and separation using nonparametric regularization method', Signal Processing 186, 2021, 108131. DOI: 10.1016/j.sigpro.2021.108131, https://doi.org/10.1016/j.sigpro.2021.108131
12. K. P. Constantinou, I. P. Constantinou, C. S. Pattichis, and M. S. Pattichis, 'Medical image analysis using AM-FM models and methods', IEEE Reviews in Biomedical Engineering 14, 2021, pp. 270-289. DOI: 10.1109/RBME.2020.2967273, https://doi.org/10.1109/rbme.2020.2967273
13. F. Salzenstein and A.-O. Boudraa, 'Multidimensional directional derivatives and AM-FM dual-band image demodulation by higher order Teager-Kaiser operators', Digital Signal Processing 129, 2022, 103641. DOI: 10.1016/j.dsp.2022.103641, https://doi.org/10.1016/j.dsp.2022.103641