Distortion-Aware Spectral Modulator (DASM)

• DASM is a dual-use technique that corrects distortion using optical phase modulation for spectral filtering and neural feature calibration for panoramic vision.
• In the optical setting, it integrates advanced pattern design with learned residual correction to boost hyperspectral imaging accuracy and material classification performance.
• In panoramic applications, DASM uses high/low-frequency decomposition and latitude-aware gating to adjust ERP feature distortions for robust affordance grounding.

Distortion-Aware Spectral Modulator (DASM) denotes two distinct technical constructs in the arXiv literature. In the computational-imaging setting of "Programmable Spectral Filter Arrays using Phase Spatial Light Modulator" (Saragadam et al., 2021), DASM is a liquid-crystal phase spatial light modulator (SLM) system whose optical model, pattern design, and learned restoration are jointly engineered to achieve high-fidelity spatially varying spectral modulation. In the panoramic-vision setting of "PanoAffordanceNet: Towards Holistic Affordance Grounding in 360° Indoor Environments" (Zhu et al., 10 Mar 2026), DASM is a fully differentiable neural module for latitude-dependent calibration of equirectangular-projection (ERP) features through frequency decomposition and gated compensation. The shared acronym therefore refers not to a single standardized method, but to two distortion-aware mechanisms developed for different signal domains, operating assumptions, and downstream tasks.

1. Terminological scope and dual usage

The term DASM is used in two non-equivalent senses. In the 2021 phase-SLM work, the accompanying technical report explicitly expands DASM as a Distortion-Aware Spectral Modulator and situates it in programmable spectral filter arrays implemented by a liquid-crystal phase SLM, with emphasis on optical aberration analysis, "good patterns" design, deep residual correction, hyperspectral imaging, material classification, and programmable filter synthesis (Saragadam et al., 2021). In the 2026 panoramic grounding work, DASM is the name of a module inserted into a vision-language network, where it performs latitude-aware distortion compensation after cross-modal semantic injection and before the spherical decoder (Zhu et al., 10 Mar 2026).

Usage	Substrate	Core function
DASM in phase-SLM imaging	Liquid-crystal phase SLM + camera	Spatially varying spectral modulation with distortion control
DASM in PanoAffordanceNet	Neural feature module	Latitude-dependent calibration of ERP features

This terminological overlap is substantive rather than merely stylistic. The optical DASM modulates measured radiance through birefringent phase retardance and polarization optics, whereas the PanoAffordanceNet DASM modulates learned feature tensors through Laplacian/Gaussian splitting, gating, and self-attention. A plausible implication is that the acronym has become a reusable label for distortion-aware modulation rather than a unique architecture.

2. Optical forward model and aberration mechanism

In the phase-SLM formulation, an SLM pixel at spatial coordinate $(x,y)$ imposes a voltage-controlled birefringent phase retardance

$$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$

where $\Delta n[v]$ is the voltage-dependent birefringence, $d_{LC}$ is the LC-cell thickness, and $\lambda$ is wavelength (Saragadam et al., 2021). Placing the SLM between crossed polarizers yields a local spectral transmittance

$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$

For a broadband scene with hyperspectral radiance $h(x,y,\lambda)$ and camera spectral response $s(\lambda)$, the measured intensity is

$$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$

After discretization into $N_\lambda$ bands, the per-frame model is written as

$$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$0

The central distortion mechanism is phase-gradient-induced aberration. Ideal constant-phase patterns, corresponding to a single LC cell, introduce no point-spread-function (PSF) distortion. By contrast, spatial gradients $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$1 locally tilt the wavefront, vignette the relay optics, and broaden the PSF. With a "gamma" curve chosen to flatten device non-linearity, the report gives the approximation

$$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$2

so large $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$3 induces strong tilt-blur. In Fourier-optics terms, the effective pupil becomes phase-modulated, and the PSF is the squared modulus of the inverse Fourier transform of that pupil. The report therefore attributes the loss of both spatial and spectral fidelity to unintended phase $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$4, which broadens the PSF and mixes neighboring spectral filters (Saragadam et al., 2021).

3. Pattern optimization and neural correction

The optical DASM addresses distortion in two stages: pattern design and learned residual correction. For pattern design, the report defines an energy $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$5 that trades off three quantities: phase-smoothness, which penalizes large spatial gradients; spectral diversity, which rewards local variation in spectral transmittance; and an implementability regularizer $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$6 that can encode tiling-periodicity or pattern constraints such as 8-bit quantization (Saragadam et al., 2021). The coefficients $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$7 determine the blur-diversity-implementability tradeoff.

Because $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$8 is non-convex and the SLM must display 8-bit patterns, optimization is performed by a two-stage strategy. First, there is discrete combinatorial search within a small local tile, exemplified by a $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$9 tile, using greedy or simulated-annealing methods to minimize $\Delta n[v]$0. Second, the best local designs are extended to the full SLM by global tiling via spatial shifts, including circular or mirror symmetries. The report also compares explicit candidate families: 1D ramps, staggered ramps, 2D periodic or mirror-symmetric tiles, and locally repeated random blocks, and then evaluates $\Delta n[v]$1 to select the top $\Delta n[v]$2 patterns (Saragadam et al., 2021).

Residual blur and vignetting are then corrected by a pattern-oblivious encoder-decoder. The learned mapping is

$\Delta n[v]$3

where $\Delta n[v]$4 is a 64-dimensional sinusoidal positional encoding of $\Delta n[v]$5. The network uses one intensity channel, optional guide-RGB channels, and positional-encoding channels; a downsampling path with $\Delta n[v]$6 followed by $\Delta n[v]$7 across four levels with channels $\Delta n[v]$8; residual blocks at the bottleneck; and an upsampling path with transposed convolutions and $\Delta n[v]$9 convolutions that reverse the channel schedule. A final linear $d_{LC}$0 convolution outputs one restored channel. Training minimizes a per-pixel $d_{LC}$1 loss against "ground-truth" simulated measurements from full-scan data, and the report notes that this simple loss suffices to recover at least $d_{LC}$2 PSNR gain over raw data (Saragadam et al., 2021).

The training corpus for this network is also specified. The prototype uses a $d_{LC}$3 SLM and a $d_{LC}$4 camera over $d_{LC}$5--$d_{LC}$6, with 42 indoor scenes under three broadband illuminants. Each scene includes a full scan of 256 constant SLM patterns and 92 spatially varying patterns comprising 1D ramps, 2D tiles, random blocks, and shifts. Patches are $d_{LC}$7, batch size is 500, optimization uses Adam with $d_{LC}$8, $d_{LC}$9, $\lambda$0, and training runs for 100k iterations on $\lambda$1 Titan Xp for roughly 45 hours. Validation on seven held-out scenes gives a typical PSNR lift of at least $\lambda$2 and an angular-error drop of at least $\lambda$3 (Saragadam et al., 2021).

4. Optical performance and application envelope

The report evaluates hyperspectral imaging, material classification, and programmable filter synthesis under a unified DASM pipeline (Saragadam et al., 2021). Hyperspectral reconstruction is posed either as a guide-free linear inverse with a data term, $\lambda$4 regularization, and a spectral smoothness term $\lambda$5, or as guided superpixel rank-1 modeling in which each superpixel is parameterized by a grayscale guide and an estimated spectrum. The stated metrics are spatial quality by PSNR in dB and line-spread-function FWHM $\lambda$6, spectral fidelity by RMSE over 53 bands and Spectral Angle Mapper (SAM) median $\lambda$7, and throughput by comparison to an LC-cell baseline with the same light-throughput of approximately $\lambda$8, while DASM adds $\lambda$9 programmability.

For single-image visible HSI on the ICVL test set, the reported mean $$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$0 standard deviation over six scenes is as follows:

Configuration	PSNR	SAM
Raw SLM (no restoration)	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$1	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$2
+ Good-Pattern selection	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$3	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$4
+ Restoration network (DASM)	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$5	$$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$6

These results are stated to outperform RGB-to-HSI priors, reported as approximately $$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$7 PSNR and approximately $$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$8 SAM, and to rival spatio-spectral coded snapshot imaging (SASSI) with guide (Saragadam et al., 2021). In the multi-image regime, the report highlights three operating points: for $$m\bigl(v(x,y),\lambda\bigr)=\tfrac12\bigl[1-\cos \phi(x,y;\lambda)\bigr].$$9, DASM with guide improves PSNR by $h(x,y,\lambda)$0 over the LC baseline; around $h(x,y,\lambda)$1, DASM without guide is approximately equal to LC while DASM with guide reaches approximately $h(x,y,\lambda)$2; and beyond $h(x,y,\lambda)$3, all methods converge to approximately $h(x,y,\lambda)$4, identified as the full-scan limit.

The non-imaging applications are equally explicit. For material classification with $h(x,y,\lambda)$5 materials, the system measures responses to 256 basis filters, chooses three SLM indices $h(x,y,\lambda)$6 that maximize the minimum simplex-distance, tiles those indices in a $h(x,y,\lambda)$7 checkerboard, and performs nearest-neighbor classification in a 3-D feature space. The reported accuracy is $h(x,y,\lambda)$8 versus $h(x,y,\lambda)$9 chance on a color-checker and plant scene. For programmable arbitrary filters, the SLM is placed in the pupil plane so that each pixel modulates the spectrum of the entire field of view; a non-negative least-squares problem

$s(\lambda)$0

is solved to approximate a target continuous filter $s(\lambda)$1. The report states that this yields bandpass Gaussians with center wavelengths $s(\lambda)$2 and $s(\lambda)$3 with simulated error below $s(\lambda)$4 RMS, while prototype spectrometer captures match within $s(\lambda)$5 (Saragadam et al., 2021).

5. DASM in panoramic affordance grounding

In PanoAffordanceNet, DASM is not an optical device but an intermediate feature-processing block positioned immediately after the dual-encoder backbone, which consists of DINOv2 visual features and CLIP text embeddings, and immediately before the spherical decoder (Zhu et al., 10 Mar 2026). Its pipeline has four stages: cross-modal semantic injection, dual-frequency decomposition into high- and low-frequency branches, branch-wise distortion compensation through HFEM and LFSM, and gated fusion followed by contextual re-aggregation. The input is visual tokens $s(\lambda)$6 and text tokens $s(\lambda)$7, and the output is a single distortion-robust, affordance-aware feature $s(\lambda)$8.

Cross-modal injection is given by

$s(\lambda)$9

with learned $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$0 and $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$1 (Zhu et al., 10 Mar 2026). The tensor is then reshaped so that $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$2, with the example $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$3, $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$4. Frequency decomposition uses a discrete Laplacian kernel

$$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$5

as a high-pass operator and a Gaussian kernel

$$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$6

as a low-pass operator, with $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$7 and kernel size $$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$8. The resulting features are

$$i(x,y)=\int_\Lambda h(x,y,\lambda)\,m\bigl(v(x,y),\lambda\bigr)\,s(\lambda)\,d\lambda.$$9

The distortion model is explicitly latitude-dependent. For each pixel $N_\lambda$0, latitude is

$N_\lambda$1

and the ERP distortion factor is proportional to $N_\lambda$2 (Zhu et al., 10 Mar 2026). DASM therefore uses a shared channel gate

$N_\lambda$3

and branch-specific spatial gates

$N_\lambda$4

where $N_\lambda$5 is a local mean or $N_\lambda$6 convolution. The high-frequency branch is compensated as

$N_\lambda$7

which the paper characterizes as up-weighting equatorial edges and suppressing polar noise, while the low-frequency branch is compensated as

$N_\lambda$8

to reinforce structural cues near poles. Gated residual fusion then forms

$N_\lambda$9

where $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$00 are learned scalars. Finally, flattening plus one Multi-Head Self-Attention layer and an FFN yields $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$01 (Zhu et al., 10 Mar 2026).

Implementation details are fully specified. The channel MLP hidden dimension is $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$02 and output dimension is $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$03; the spatial MLP hidden dimension is 64 with scalar output per branch; LoRA rank is $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$04 in the backbone attention layers; the MHSA uses 8 heads and the FFN hidden dimension is $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$05. DASM has no dedicated auxiliary loss; its output flows into $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$06, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$07, and $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$08, with final loss

$$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$09

and default weights $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$10. Training freezes most of DINOv2 and CLIP, updating only LoRA adapters, DASM modules, decoder, and loss heads, using AdamW with learning rate $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$11, weight decay $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$12, cosine annealing, panoramic data augmentation by horizontal wrap, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$13 rotation, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$14 scale, and color jitter, with batch size 4 on two A6000 GPUs for 20k iterations (Zhu et al., 10 Mar 2026).

The empirical ablation reported for the Hard split isolates DASM's contribution. LoRA only gives $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$15, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$16, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$17; adding DASM without OSDH gives $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$18, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$19, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$20; adding OSDH without DASM gives $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$21, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$22, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$23; and the full system gives $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$24, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$25, $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$26 (Zhu et al., 10 Mar 2026). The paper interprets DASM alone as reducing KLD by about $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$27 and improving SIM and NSS, and qualitatively reports that methods without distortion-aware spectral calibration show fragmented, pole-biased activations, whereas DASM yields crisp, geographically consistent affordance maps across the full $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$28 field.

6. Conceptual relation, significance, and recurrent points of confusion

The two DASMs are connected by a common design principle but not by a common mechanism. The optical DASM models physically induced distortion arising from unintended phase gradients in a phase SLM and mitigates it through pattern smoothness constraints plus learned restoration (Saragadam et al., 2021). The PanoAffordanceNet DASM models geometric distortion arising from ERP latitude effects and mitigates it through explicit high/low-frequency decomposition, latitude-aware gating, and self-attentive re-aggregation (Zhu et al., 10 Mar 2026). In both cases, distortion is treated as structured rather than incidental, and correction is embedded directly into the modulation pipeline.

A recurring source of confusion is the word spectral. In the phase-SLM work, spectral modulation is literal optical wavelength modulation: the transmittance $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$29 acts on scene radiance as a function of $$\phi(x,y;\lambda)=2\pi\,\Delta n[v(x,y)]\cdot d_{LC}/\lambda,$$30 (Saragadam et al., 2021). In PanoAffordanceNet, the same term appears in a signal-processing sense tied to high- and low-frequency branches constructed by Laplacian and Gaussian filtering, not to optical wavelength selection (Zhu et al., 10 Mar 2026). This distinction matters because identical terminology can suggest continuity of hardware, whereas the later DASM is entirely feature-domain.

The significance of the original DASM lies in demonstrating that careful phase-gradient analysis, "good patterns" selection, and a restoration network can recover high-fidelity, high-resolution spatio-spectral modulation with applications spanning dynamic spectral filtering, hyperspectral imaging, material classification, and programmable filter synthesis (Saragadam et al., 2021). The significance of the later DASM lies in showing that a lightweight, fully differentiable module with explicit latitude-aware gating can improve one-shot affordance grounding in panoramic scenes by reducing ERP-induced representation bias (Zhu et al., 10 Mar 2026). This suggests a broader methodological pattern: distortion-aware modulation can be formulated either as a physical sensing strategy or as an internal representation strategy, provided the distortion mechanism is made explicit and computationally tractable.