Robust multiscale AM-FM demodulation and image reconstruction for real digital images: out-of-passband and low instantaneous frequencies, nonideal sampling, and principled filterbank selection

Multiscale amplitude-modulation frequency-modulation (AM-FM) demodulation represents a 2-D image as a sum of components $a_n(x,y)\cos\varphi_n(x,y)$ and estimates per-pixel instantaneous amplitude (IA) and instantaneous frequency (IF) $\nabla\varphi_n$ through a bank of bandpass channel filters. The framework of Murray, Rodriguez and Pattichis (IEEE TIP, 2010) achieved large accuracy gains with a flat-passband equiripple filterbank, a condition-number-managed local-linear-phase IF estimator, a full-band extension of the quasi-local method, and least-squares reconstruction from few AM-FM basis functions — but it was developed and validated under idealizing assumptions: perfectly bandlimited, ideally sampled signals, an instantaneous frequency lying inside some channel passband at almost every pixel, sufficient AM/FM spectral separation, and manually chosen filterbank bandwidths. This problem asks for a demodulation-and-reconstruction framework for real discrete images that removes those assumptions: it must detect and accurately estimate IFs lying outside every channel passband without oracle band knowledge, preserve accuracy at very low IF magnitudes where AM and FM spectra overlap, remain principled under nonideal sampling and aliasing, select the filterbank and channels by a data-driven rule, and support reconstruction from small AM-FM bases that is free of filterbank- and reconstruction-induced artifacts. The estimation-theoretic core of these requirements remains open in the literature.

Background

An AM-FM image model expands an input image as $I(x,y)=\sum_{n=1}^{M}a_n(x,y)\cos\varphi_n(x,y)$, where the instantaneous amplitude (IA) functions $a_n(x,y)$ describe slowly varying brightness or contrast and the instantaneous phase functions $\varphi_n(x,y)$ describe spatially varying oscillations; the instantaneous frequency (IF) is the gradient field $\nabla\varphi_n(x,y)$, measured in radians per sample. Demodulation means recovering IA, phase, and IF per pixel. Unlike Fourier or wavelet decompositions with fixed bases, AM-FM components are nonstationary, so they capture locally coherent texture such as retinal lesions, ultrasound speckle patterns, and fingerprints. Classical demodulation routes include the energy separation algorithm (ESA) built on Teager–Kaiser energy operators (nonlinear differential operators whose discrete form estimates $|\nabla\varphi|^2$ and $a^2$ from local samples), the quasi-eigenfunction approximation (QEA, which treats a filtered AM-FM component as a local eigenfunction of the channel filter so that IA/IF can be read from the filter output), quasi-local methods (QLM, algebraic estimators combining several nearby samples of the analytic image through inverse trigonometric formulas), and multicomponent pipelines built on dominant component analysis (DCA), which assigns each pixel to the single filterbank channel with the largest response and takes that channel's estimate. To apply QEA or QLM the image is prefiltered by a separable bank of bandpass filters covering the discrete frequency plane; the analytic image is formed with the partial Hilbert transform (a 2-D extension of the Hilbert transform that shifts one spatial frequency axis by 90 degrees). Murray, Rodriguez and Pattichis (2010) advanced this machinery on four fronts: (1) a multiscale filterbank of 1-D min-max (equiripple) filters with near-unity flat passbands (0.017 dB ripple, 66 dB stopband attenuation), so QEA amplitude estimates no longer need correction by the filter gain at the estimated IF; (2) a variable-spacing local linear phase (VS-LLP) IF estimator that picks sample spacing and inverse trigonometric function by condition number $|xf'(x)/f(x)|$ to avoid numerically ill-conditioned evaluations; (3) a two-filter extension of QLM covering the whole discrete frequency range $[0,\pi]$ via the substitution $\xi=\pi-\omega$, avoiding resampling; and (4) least-squares reconstructions (LESHA, MULTILES, LESHAL) that combine a DC term, the lowpass channel, and a handful of AM-FM components or harmonics. The development, however, rests on assumptions that real digital images violate. First, the derivations assume ideal continuous-space, bandlimited signals, while real images are sampled, not necessarily bandlimited, and aliased. Second, estimation is only valid at pixels whose IF falls within the passband of a dominant channel; the paper's modulation trick handles IFs in $[\pi/2,\pi]$ only when the correct band is identifiable, and pixels whose IF lies outside every passband (or straddles band edges) are mishandled with no detection rule. Third, estimation accuracy degrades at very low IF magnitudes, because the AM spectrum concentrates near zero frequency and overlaps the FM component's spectrum, breaking the separation the estimators need. Fourth, the per-filter passband bandwidths, band overlaps, and scale counts must be hand-tuned; the authors themselves report that this choice dramatically affects accuracy, and no automated or optimal selection procedure was given. Fifth, reconstructions from few AM-FM basis functions show visible artifacts on structures such as Lena's shoulder, indicating the small bases misrepresent certain image content. The same group and others subsequently built many applications (diabetic retinopathy screening, MRI, ultrasound, texture classification) on this framework, so the reliability limits of the underlying estimates propagate directly into those applications.

Problem Statement

Construct, analyze, and validate a multiscale AM-FM demodulation-and-reconstruction framework for real discrete digital images — i.e., an image model $I(x,y)=\sum_n a_n(x,y)\cos\varphi_n(x,y)$ together with a filterbank of bandpass channels and per-pixel estimators of instantaneous amplitude $a_n$, phase $\varphi_n$, and instantaneous frequency $\nabla\varphi_n$, plus a reconstruction scheme from a small number of AM-FM basis functions — that removes the idealizing assumptions under which the Murray-Rodriguez-Pattichis (2010) framework was developed. These five requirements are required jointly and serve one objective: obtaining dependable per-pixel IA/IF estimates and artifact-limited reconstructions on real images, because each listed failure mode corrupts the same estimates that every downstream use consumes. Specifically, the framework must: (1) detect and accurately handle pixels whose instantaneous frequency lies outside the passband of every channel filter — including frequencies near or beyond Nyquist band edges and components whose band membership is not identifiable from a dominant response — without oracle knowledge of the correct band and beyond the fixed $\xi=\pi-\omega$ modulation trick; (2) preserve IA/IF estimation accuracy at very low IF magnitudes, where the amplitude spectrum overlaps the frequency-modulation spectrum, without requiring clear AM/FM spectral separation; (3) remain principled under nonideal sampling — sampling, aliasing, and reconstruction error for real digital images that are not bandlimited continuous-space signals — through error bounds or bias/variance analyses for discrete nonstationary estimators; (4) select the filterbank itself (per-filter passband bandwidths, band overlap, number of scales and channels, and per-pixel channel assignment) by a principled, data-driven or adaptive rule in place of hand tuning; and (5) support reconstructions from few AM-FM components that do not introduce filterbank- or reconstruction-induced artifacts — such as ringing/fringe patterns and boundary discontinuities — on image structures poorly represented by small AM-FM bases, as distinct from the unavoidable representation error of a small AM-FM basis. Accuracy claims must be quantified against ground-truth synthetic AM-FM images spanning the stress regimes named above (near-zero IF, out-of-passband IF, non-bandlimited or aliased sampling) and on natural images, in direct comparison with the 2010 framework's QEA/QLM/VS-LLP estimators and least-squares reconstructions.

References

1. Murray, V.; Rodriguez, P.; Pattichis, M. S. (2010). Multiscale AM-FM Demodulation and Image Reconstruction Methods With Improved Accuracy. IEEE Transactions on Image Processing 19(5):1138-1152. DOI: 10.1109/TIP.2010.2040446. https://doi.org/10.1109/TIP.2010.2040446 (source paper)
2. Murray, V.; Pattichis, M. S.; Barriga, E. S.; Soliz, P. (2012). Recent multiscale AM-FM methods in emerging applications in medical imaging. EURASIP Journal on Advances in Signal Processing 2012:23. DOI: 10.1186/1687-6180-2012-23. https://doi.org/10.1186/1687-6180-2012-23 (open access)
3. Murray, V.; Pattichis, M. S.; Soliz, P. (2011). Multiscale directional AM-FM demodulation of images using a 2D optimized method. 18th IEEE International Conference on Image Processing, pp. 249-252. DOI: 10.1109/ICIP.2011.6116169. https://doi.org/10.1109/ICIP.2011.6116169
4. Constantinou, I.; Pattichis, M. S.; Pattichis, C. S. (2013). Multiscale AM-FM image reconstructions based on elastic net regression and Gabor filterbanks. 2013 Asilomar Conference on Signals, Systems and Computers, pp. 1985-1989. DOI: 10.1109/ACSSC.2013.6810653. https://doi.org/10.1109/ACSSC.2013.6810653
5. Constantinou, K. P.; Constantinou, I. P.; Pattichis, C. S.; Pattichis, M. S. (2021). Medical Image Analysis Using AM-FM Models and Methods. IEEE Reviews in Biomedical Engineering 14:270-289. DOI: 10.1109/RBME.2020.2967273. https://doi.org/10.1109/RBME.2020.2967273
6. Liu, W.; Santhanam, B. (2016). Wideband image demodulation via bi-dimensional multirate frequency transformations. Journal of the Optical Society of America A 33(9):1668. DOI: 10.1364/JOSAA.33.001668. https://doi.org/10.1364/JOSAA.33.001668
7. Ulloa, A.; Rodriguez, P.; Liu, J.; Calhoun, V.; Pattichis, M. (2014). A quasi-local method for instantaneous frequency estimation with application to structural magnetic resonance images. 36th Annual International Conference of the IEEE EMBC, pp. 1477-1480. DOI: 10.1109/EMBC.2014.6943880. https://doi.org/10.1109/EMBC.2014.6943880
8. Carranza, C.; Murray, V.; Pattichis, M.; Barriga, E. S. (2012). Multiscale AM-FM decompositions with GPU acceleration for diabetic retinopathy screening. 2012 IEEE Southwest Symposium on Image Analysis and Interpretation, pp. 121-124. DOI: 10.1109/SSIAI.2012.6202468. https://doi.org/10.1109/SSIAI.2012.6202468
9. Llamocca, D.; Carranza, C.; Pattichis, M. (2012). Dynamic multiobjective optimization management of the Energy-Performance-Accuracy space for separable 2-D complex filters. 22nd International Conference on Field Programmable Logic and Applications, pp. 579-582. DOI: 10.1109/FPL.2012.6339211. https://doi.org/10.1109/FPL.2012.6339211
10. Llamocca, D.; Pattichis, M. (2014). A Self-Reconfigurable Platform for the Implementation of 2D Filterbanks with Real and Complex-Valued Inputs, Outputs, and Filter Coefficients. VLSI Design 2014:651943, pp. 1-24. DOI: 10.1155/2014/651943. https://doi.org/10.1155/2014/651943
11. Murray, V.; Agurto, C.; Barriga, S.; Pattichis, M. S.; Soliz, P. (2012). Real-time diabetic retinopathy patient screening using multiscale AM-FM methods. 19th IEEE International Conference on Image Processing, pp. 525-528. DOI: 10.1109/ICIP.2012.6466912. https://doi.org/10.1109/ICIP.2012.6466912
12. Agurto, C.; Murray, V.; Barriga, E.; Murillo, S.; Pattichis, M.; Davis, H.; Russell, S.; Abramoff, M.; Soliz, P. (2010). Multiscale AM-FM Methods for Diabetic Retinopathy Lesion Detection. IEEE Transactions on Medical Imaging 29(2):502-512. DOI: 10.1109/TMI.2009.2037146. https://doi.org/10.1109/TMI.2009.2037146
13. Clausel, M.; Oberlin, T.; Perrier, V. (2012). The Monogenic Synchrosqueezed Wavelet Transform: A tool for the Decomposition/Demodulation of AM-FM images. arXiv:1211.5082. https://arxiv.org/abs/1211.5082
14. Sebghati, M.; Amindavar, H.; Ritcey, J. A. (2014). Basis approach to estimate the instantaneous frequencies in multicomponent AM-FM signals. EURASIP Journal on Audio, Speech, and Music Processing 2014:8. DOI: 10.1186/1687-4722-2014-8. https://doi.org/10.1186/1687-4722-2014-8
15. Salzenstein, F.; Boudraa, A.-O. (2022). Multidimensional directional derivatives and AM-FM dual-band image demodulation by higher order Teager-Kaiser operators. Digital Signal Processing 129:103641. DOI: 10.1016/j.dsp.2022.103641. https://doi.org/10.1016/j.dsp.2022.103641
16. Constantinou, K. P.; Constantinou, I. P.; Pattichis, M. S.; Kyriacou, E.; Pattichis, C. S. (2023). Carotid plaque stroke risk assessment using multiscale AM-FM analysis based on DoG filterbanks. 2023 24th International Conference on Digital Signal Processing, pp. 1-4. DOI: 10.1109/DSP58604.2023.10167874. https://doi.org/10.1109/DSP58604.2023.10167874
17. Shi, W. (2022). Human Attention Detection Using AM-FM Representations. Master's thesis. arXiv:2203.07093. https://arxiv.org/abs/2203.07093