Dual Modulation Framework Overview

• Dual Modulation Framework is a family of techniques that couple two interacting structures, such as dual task heads or attention and fusion mechanisms, to enhance performance.
• It spans diverse applications including RF signal classification, UHD image restoration, and RGB-T crowd counting, achieving significant accuracy and efficiency gains.
• Approaches vary from joint multi-task learning and dual-domain decoupling to index modulation and duality-based control, reflecting its methodological versatility and domain-specific implementations.

“Dual Modulation Framework” is not a single standardized construct. In current literature, the label denotes several technically distinct schemes that couple two mechanisms, domains, or tasks within one architecture. In wireless signal recognition, it denotes joint automatic modulation classification and signal-type recognition via multi-task learning (Jagannath et al., 2021). In RGB-T crowd counting, it denotes the combination of Spatially Modulated Attention and Adaptive Fusion Modulation (Feng et al., 21 Sep 2025). In UHD image restoration, it denotes dual-domain decoupled spectral modulation, with explicit amplitude-spectrum modulation and spatial-domain refinement (Zhao et al., 7 Nov 2025). In communications, closely related uses include joint spectral-spatial index modulation, dual-mode chirp spread spectrum, dual-polarization nonlinear spectral modulation, and RF-optical cross-band mapping (Huang et al., 2019, Azim et al., 2022, Ros et al., 2019, Oikonomou et al., 21 Mar 2025). In control theory, it denotes dual modulating functions and dual modulations derived from auxiliary systems and duality (Accioli et al., 5 Jul 2026).

1. Terminological scope and domain-specific meanings

The expression is therefore best treated as a family of domain-specific formulations rather than as a universal algorithm. A recurrent pattern is the coupling of two interacting representational spaces, but the coupled objects differ sharply: task heads in RF classification, attention and fusion operators in vision, amplitude and phase handling in restoration, spectrum and spatial indices in waveform design, or primal and dual operators in control and optimization.

Usage domain	Coupled elements	Representative source
Wireless signal recognition	AMC and signal-type recognition	(Jagannath et al., 2021)
RGB-T crowd counting	SMA and AFM	(Feng et al., 21 Sep 2025)
UHD restoration	Frequency-domain amplitude and spatial-domain phase-related refinement	(Zhao et al., 7 Nov 2025)
Semantic segmentation DG	Stage-1 style perturbation and Stage-2 pixel-wise blending	(Tjio et al., 2023)
RF/optical and radar communications	Two signaling dimensions or bands	(Huang et al., 2019, Azim et al., 2022, Oikonomou et al., 21 Mar 2025)
Control and optimization	Modulation operators and dual variables	(Accioli et al., 5 Jul 2026, Ahmad et al., 2011)

A common misconception is that “dual modulation” always refers to physical-layer symbol mapping. The corpus shows otherwise. In some works, modulation is literal waveform or constellation design; in others, it is feature reweighting, stylization control, or an operator-theoretic construction. This heterogeneity is central to the term’s current scholarly usage.

2. Joint-task and dual-driven inference in RF sensing

In wireless signal recognition, the framework in “Multi-task Learning Approach for Automatic Modulation and Wireless Signal Classification” formulates a dual-task problem: modulation classification over 9 classes and signal-type classification over 11 classes from the same RF input (Jagannath et al., 2021). Each example contains 128 complex IQ samples at 10 MS/s, normalized to unit energy, real-vectorized from $\mathbb{C}^{128\times1}$ to $\mathbb{R}^{256\times1}$, and reshaped to a $16\times16$ tensor. The network uses hard parameter sharing, with a shared 2D convolutional trunk and two task-specific heads. The tuned lightweight configuration uses 8 shared filters, 4 filters in each head, and 256-unit dense layers; kernel size $3\times3$ yields the fastest convergence. The joint loss is

$$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$

with categorical cross-entropy on both heads and the adopted weights $w_m=0.2$, $w_s=0.8$. On RadComAWGN, the reported modulation accuracies are 97.87% at 0 dB, 98.58% at 2 dB, and 99.53% at 10 dB; signal-type accuracies are 92.3% at 0 dB, 97.87% at 2 dB, and 99.53% at 10 dB. On RadComDynamic, after transfer learning from RadComAWGN, the model reaches 97.07% modulation accuracy and 90.86% signal-type accuracy at 2 dB (Jagannath et al., 2021).

A related but distinct formulation appears in “Data-and-Knowledge Dual-Driven Automatic Modulation Recognition for Wireless Communication Networks,” where the duality is between a visual model and an attribute learning model, connected through a transformation model that maps attribute representations into the visual feature space (Ding et al., 2022). The visual branch uses raw $2\times128$ IQ input and a Multi-Scale Convolutional Network, while the attribute branch predicts a 6-dimensional semantic attribute vector. Training is staged: the visual model is pretrained with cross-entropy, the attribute model with mean squared error, and the transformation network with a least-squares consistency loss. Reported results show approximately 60% accuracy at $-20$ dB, more than 70% at $-14$ dB, and at 12 dB gains of +2% over ResNet, +5% over MSNet, and +17% over LSTM and VGG; the framework also reduces confusion between 16QAM and 64QAM (Ding et al., 2022).

The same joint-design logic extends from classification to detection. In “Hierarchical Feature Integration for Multi-Signal Automatic Modulation Recognition,” dual-modulation is treated as a special case of multi-signal spectrogram detection and AMR, with coexisting or overlapping RF signals localized in time-frequency space and labeled by modulation class (Qu et al., 28 Nov 2025). HIFI-YOLO operates on STFT spectrograms and integrates low-frequency enhancement, content-adaptive resampling, and decoupled prediction heads. On a 36,000-sample dataset with 14 modulation types and 3–5 coexisting signals per sample, the method reports $\mathbb{R}^{256\times1}$0, $\mathbb{R}^{256\times1}$1, and $\mathbb{R}^{256\times1}$2, outperforming YOLOv5-S and other detectors (Qu et al., 28 Nov 2025). The paper explicitly notes that dual-modulation with $\mathbb{R}^{256\times1}$3 is handled by the same framework without architectural change.

3. Feature modulation in multimodal vision and restoration

In RGB-T crowd counting, the “Dual Modulation Framework” consists of Spatially Modulated Attention and Adaptive Fusion Modulation (Feng et al., 21 Sep 2025). SMA inserts a learnable Spatial Decay Mask into self-attention,

$\mathbb{R}^{256\times1}$4

where $\mathbb{R}^{256\times1}$5 is derived from pairwise Euclidean token distances using per-head learnable parameters $\mathbb{R}^{256\times1}$6 and $\mathbb{R}^{256\times1}$7. AFM computes a global fusion weight $\mathbb{R}^{256\times1}$8 from $\mathbb{R}^{256\times1}$9 and fuses modality features as

$16\times16$0

The architecture uses a weight-sharing VGG-19 backbone, two 2-layer 8-head Transformer encoders, and a regression head. On RGBT-CC, the model reports $16\times16$1, $16\times16$2, $16\times16$3, $16\times16$4, and $16\times16$5; on DroneRGBT it reports $16\times16$6 and $16\times16$7. Ablations attribute a reduction from $16\times16$8 to 11.09 and from $16\times16$9 to 20.53 to SMA alone, with AFM further improving $3\times3$0 from 11.09 to 10.80 and $3\times3$1 from 20.53 to 19.87 (Feng et al., 21 Sep 2025).

In UHD image restoration, “UHDRes: Ultra-High-Definition Image Restoration via Dual-Domain Decoupled Spectral Modulation” defines dual modulation as asymmetric use of spatial and frequency domains (Zhao et al., 7 Nov 2025). The frequency branch explicitly modulates the amplitude spectrum,

$3\times3$2

while phase is left untouched in the spectrum domain and refined implicitly by spatial operations. The spatio-spectral fusion module combines a Multi-Scale Context Aggregator with a Decoupled Spectral Modulation Block, and the Shared Gated Feed-Forward Network uses shared strip convolutions for low-cost directional context. The total loss is

$3\times3$3

with $3\times3$4. The model has approximately 0.40M parameters and reports, among other results, PSNR 27.693 on UHD-LL, 25.575 on UHD-Haze, 31.480 on 8KDehaze-mini, 31.453 on UHD-Blur, and 34.934 on 4K-Rain13k (Zhao et al., 7 Nov 2025). An important negative result is that modulating amplitude only is best; simultaneously modulating phase worsens performance.

A different vision usage appears in “Dual Stage Stylization Modulation for Domain Generalized Semantic Segmentation,” where ASH+ performs dual-stage manipulation of hallucination strength (Tjio et al., 2023). Stage 1 perturbs style features through affine parameters $3\times3$5 and $3\times3$6 produced by the dFT layer; Stage 2 blends content and AdaIN-transformed features using a learned per-pixel weight $3\times3$7:

$3\times3$8

The adversarial interplay is realized without a discriminator by maximizing a pixel-wise consistency loss with respect to dFT while minimizing it with respect to the segmentation backbone. On SYNTHIA and Cityscapes, the reported gains over source-only and prior stylization baselines are attributed to semantic-aware spatial control of stylization, especially for minority classes (Tjio et al., 2023).

4. Dual modulation in communication signaling and joint spectrum usage

In communications, the phrase often denotes literal symbol, spectrum, or index modulation. “MAJoRCom: A Dual-Function Radar Communication System Using Index Modulation” embeds information simultaneously in frequency selection and antenna allocation within CAESAR radar (Huang et al., 2019). Each pulse selects $3\times3$9 carriers from $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$0 and partitions $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$1 antennas into $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$2 subarrays, yielding

$$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$3

bits per pulse, with the large-system approximation $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$4. The design preserves radar performance because communication is embedded in indices rather than in amplitude/phase perturbations of the radar tone. The paper also derives a low-complexity decoder and reports that MAJoRCom can achieve a bit rate comparable to independent communication modules without affecting radar performance (Huang et al., 2019).

“Dual-Mode Chirp Spread Spectrum Modulation” expands the CSS symbol alphabet by simultaneously activating one even index and one odd index, assigning binary phase states to both, and choosing chirp direction (Azim et al., 2022). With $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$5, DM-CSS carries

$$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$6

bits per symbol, whereas LoRa carries $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$7. The reported spectral-efficiency increase relative to LoRa is therefore up to 116.66% at $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$8, and 108.33% at $$\mathcal{L}_{\mathrm{mtl}} = w_m \mathcal{L}_m + w_s \mathcal{L}_s,$$9. The receiver remains non-coherent, using dual dechirping, two FFTs, parity-restricted peak searches, and polarity tests. The paper further reports approximately 0.6 dB penalty at phase offset $w_m=0.2$0 and approximately 0.8 dB penalty at normalized CFO $w_m=0.2$1 at BER $w_m=0.2$2 (Azim et al., 2022).

Optical communications supply two further variants. “Dual-polarization NFDM transmission with continuous and discrete spectral modulation” jointly modulates the continuous and discrete nonlinear spectra across both polarizations of the Manakov system (Ros et al., 2019). The transmitter maps 64 bits per NFDM symbol to the continuous spectrum and 8 bits to the discrete spectrum, for 72 bits per symbol, and experimentally demonstrates 8.4 Gb/s net throughput up to 3200 km. “Extending fibre nonlinear interference power modelling to account for general dual-polarisation 4D modulation formats” is not a signaling proposal but a modeling framework: it extends EGN analysis from PM-2D to arbitrary dual-polarisation 4D formats and shows that NLI power depends not only on intra-polarisation high-order moments but also on cross-polarisation correlations (Liga et al., 2020).

Hybrid physical-layer amalgamation appears in optical wireless and RF-optical systems. “A Framework for Amalgamating Optical-OFDM and Optical-OFDM with Index Modulation for Optical Wireless Systems” defines a hybrid symbol carrying

$w_m=0.2$3

and, for HO-OFDM-IM with $w_m=0.2$4, $w_m=0.2$5, and $w_m=0.2$6, reports peak spectral efficiency of approximately 2.063 bits/s/Hz. At approximately 2 bits/s/Hz and BER $w_m=0.2$7, the required $w_m=0.2$8 values are 18.04 dB for HO-OFDM-IM, 25.88 dB for ACEO-OFDM-IM, and 22.58 dB for ACO-OFDM-IM (Azim et al., 2022). “Cross-Band Modulation Design for Hybrid RF-Optical Systems” forms a 3D constellation in I–Q–optical-power space, with a linear mapping

$w_m=0.2$9

and an $w_s=0.8$0 receiver based on projection and rounding; a DNN-generated mapping further improves performance and is reported to deliver more than 2 dB SEP gains over the optimized linear mapping (Oikonomou et al., 21 Mar 2025).

5. Duality-based formulations in control, scheduling, and 6G channel analysis

In uplink SC-FDMA, the dual notion is explicitly optimization-theoretic. “Canonical Dual Method for Resource Allocation and Adaptive Modulation in Uplink SC-FDMA Systems” transforms binary-integer resource-allocation and adaptive-modulation problems into continuous-space canonical dual problems that are concave maximization problems (Ahmad et al., 2011). The original SUmax and JAMSCmin formulations encode contiguity through pattern matrices and binary variables, while the canonical dual transformation yields dual objectives with provable zero duality gap when the dual solution lies in the open positive cone. The framework thereby integrates contiguous resource allocation and modulation selection under SC-FDMA-specific equal-power and adjacency constraints (Ahmad et al., 2011).

In control theory, “On the Modulating Function Method for Control Problems” introduces dual modulating functions for SISO input–output systems and dual modulations for LTV state-space systems (Accioli et al., 5 Jul 2026). The central auxiliary system is

$w_s=0.8$1

and the existence of an input solving

$w_s=0.8$2

is characterized by the necessary and sufficient condition $w_s=0.8$3. Within this framework, full-state feedback, output feedback, LTI sliding mode control, the reachability Gramian, and the state-transition matrix arise as particular cases. The paper also proposes a fixed-time modulating-function regulator and reports numerical performance comparable to an LQR benchmark even under unmatched disturbances (Accioli et al., 5 Jul 2026).

A related mathematical abstraction appears in “Equivalence of Several 6G Modulation Schemes for Doubly-Selective Channels,” which analyzes modulation schemes through two properties: non-selectivity and predictability (Mehrotra et al., 12 Nov 2025). The paper shows that delay-Doppler, chirp, and time-sequency modulations are non-selective, predictable, and equivalent to one another, whereas time-frequency modulations are selective and non-predictable. This is a different use of modulation from the feature-gating or spectrum-allocation senses above, but it still relies on a paired conceptual structure: diversity is tied to non-selectivity and spectral efficiency to predictability (Mehrotra et al., 12 Nov 2025).

6. Comparative themes, misconceptions, and open problems

Across these uses, a plausible synthesis is that “dual modulation” consistently denotes a controlled interaction between two informative structures rather than simple duplication. In RF recognition, the paired structures are two labels on the same IQ tensor (Jagannath et al., 2021). In RGB-T crowd counting, they are local spatial attenuation and global scene-level modality weighting (Feng et al., 21 Sep 2025). In UHD restoration, they are explicit amplitude processing and implicit phase-related recovery (Zhao et al., 7 Nov 2025). In MAJoRCom and DM-CSS, they are distinct signaling degrees of freedom embedded in one waveform (Huang et al., 2019, Azim et al., 2022).

This also clarifies why the term does not identify a single transferable blueprint. Some frameworks are end-to-end discriminative, some are generative or restoration-oriented, some are waveform-design schemes, and some are duality-based mathematical programs. The shared vocabulary should not be mistaken for shared implementation details.

The open problems are likewise domain-specific. In wireless signal recognition, generalization to unseen signal types, open-set recognition, and domain shift across channel conditions remain unresolved (Jagannath et al., 2021). In RGB-T crowd counting, a global scalar fusion gate may underperform when modality reliability varies spatially, and thermal-RGB misregistration can undermine distance-based masking (Feng et al., 21 Sep 2025). In UHD restoration, phase sensitivity, fixed FFT resolution after pooling, and combined degradations remain explicit limitations (Zhao et al., 7 Nov 2025). In ASH+, sensitivity to $w_s=0.8$4 and $w_s=0.8$5 and dependence on early semantic predictions remain open issues (Tjio et al., 2023). In multi-signal AMR, extremely low SNR, near-identical spectral footprints, and NMS merging are identified failure cases (Qu et al., 28 Nov 2025). In dual-polarization NFDM, the chief current limitation is numerical precision in NFT/INFT processing (Ros et al., 2019). In the modulating-function regulator, the paper reports numerical robustness but does not provide formal ISS or fixed-time robustness theorems (Accioli et al., 5 Jul 2026).

For encyclopedic purposes, the most accurate definition is therefore relational rather than taxonomic: a Dual Modulation Framework is any framework in which performance is improved by coupling two modulation mechanisms, two representational domains, or two task spaces in a single end-to-end design. The literature shows that this coupling can be implemented through hard parameter sharing, spatial attention masks, global fusion gates, spectral-spatial decoupling, index modulation across physical dimensions, or duality-driven operator constructions. The term’s technical meaning is thus determined entirely by the surrounding field-specific formalism rather than by a single cross-domain canonical model.