Parabolic Directional Signatures for Lightweight All-in-One Restoration

A single image-restoration network that handles noise, rain and haze without being told which one it faces must first work out what went wrong with the picture. Current "all-in-one" restorers learn this diagnosis with extra trainable machinery: a contrastive encoder in AirNet, or a prompt generator in PromptIR that averages features over the whole image. Both sit on top of heavy backbones. We propose instead to diagnose the corruption with a fixed, parameter-free measuring instrument built from parabolic scaling. Parabolic scaling is the harmonic-analysis rule that filters should become needle-like at fine scales, with width proportional to length squared, so that they line up with curves and streaks. Our instrument reports how image energy is spread across scales and orientations, and that short "directional signature" drives a small, activation-free restoration network suitable for phones.

The intuition is that the main corruptions leave very different fingerprints in a scale-by-orientation energy map. Sensor noise adds energy that is flat across orientations and rises at fine scales. Rain adds a sharp orientation peak at middle-to-fine scales. Motion blur removes fine-scale energy in one direction. Haze flattens contrast in the coarse bands. Needle-like filters resolve these differences more sharply than the three-orientation wavelets used in today's efficient restorers. The signature can also be computed by streaming small spatial filters and pooling the outputs, so the 8–48× memory cost of storing a full shearlet decomposition never arises. The bet is that a fixed signature conditions a small network as well as a learned prompt does on the training degradations, and better on unseen levels, unseen datasets and mixtures, at a fraction of the parameters and latency.

Core Claim

Our central claim is narrow. For degradation-aware conditioning in a sub-5M-parameter restorer, a parameter-free signature computed with parabolically scaled directional filters can replace a learned degradation estimator. Doing so preserves in-distribution accuracy and improves robustness out of distribution. The parabolic aspect ratio specifically matters: an isotropic multi-orientation bank with the same number of filters should do measurably worse. Parabolic scaling is used here as the measuring instrument, not as the feature representation of the backbone. That is what separates this proposal from simply swapping the wavelet layer of an efficient network for shearlets.

Concretely, let $y$ be the degraded image and $\psi_{j,k}$ a compactly supported spatial filter at scale $j$ and orientation $k$. Its support has width about $2^{-j}$ and length about $2^{-j/2}$, and the number of orientations doubles every other scale, following the cone-adapted construction of Construction of Compactly Supported Shearlet Frames. The signature over an image tile $\Omega$ is

$$ e_{j,k}(\Omega) = \frac{\sum_{p \in \Omega} |(y * \psi_{j,k})(p)|^2}{\sum_{j',k'} \sum_{p \in \Omega} |(y * \psi_{j',k'})(p)|^2}, $$

that is, the fraction of tile energy falling in each scale–orientation cell, together with a low-pass contrast statistic. With four scales and 4, 4, 8 and 8 orientations, the signature has about 25 numbers per tile. Computing it on a coarse grid of tiles lets the conditioning vary over the image, which matters for spatially uneven haze and rain.

Relation to Prior Work

Two lines of all-in-one work motivate the gap. All-In-One Image Restoration for Unknown Corruption (AirNet) learns a degradation embedding by contrastive training and injects it through deformable convolutions. PromptIR predicts image-specific prompts from globally pooled features inside a Restormer backbone and beats AirNet by 0.86 dB on average. Both spend learned capacity on diagnosis and run on backbones that are expensive on edge hardware. Global pooling also discards how energy is split across orientations, the very information that separates rain from noise. We keep the conditioning idea and replace the learned estimator with a fixed one.

The method is supplied by work on fixed directional operators. Revisiting non-learned operators based deep learning for image classification (LDAN) shows that a non-learned shearlet bank feeding a shallow network classifies images with 80.83% fewer parameters than a wavelet hybrid, with feature sparsity 0.5671 versus nearly 1.0 for wavelets. Recognizing the corruption is itself a classification problem, so LDAN is the closest precedent. However, LDAN recognizes object categories, never restores pixels, and never conditions anything. Learning The Invisible uses shearlet coefficients to decide which parts of a CT signal a network must learn. We borrow the spirit of letting the transform route learning, but apply it across natural-image corruptions rather than one known operator.

The constraints come from edge work. tfShearlab and ShearLab 3D realize shearlets with full-image FFTs and highly redundant outputs, which is unsuitable for mobile chips. We therefore use short spatial filters on a downsampled pyramid and pool the outputs immediately. The backbone follows Simple Baselines for Image Restoration (NAFNet), which removes nonlinear activations. Its reported 40.30 dB on SIDD and 33.69 dB on GoPro show such backbones are strong single-task restorers, though it has no mechanism for blind multi-task use.

The efficient wavelet restorers EWT and PW-FNet show that transform frontends cut cost, but they use isotropic wavelets and PW-FNet trains a separate model per task. We do not claim to beat them as single-task models. Instead we use PW-FNet's isotropic wavelet sub-bands as a natural ablation of the signature.

Research Program

We will build three components.

The first is the signature extractor. We will design a digital parabolic filter bank of short spatial kernels, roughly 7×7 to 15×15 at the finest scale, applied to a two-to-three-level image pyramid. The kernels will be obtained by truncating compactly supported shearlet generators and then refitting them to be separable or low-rank, so that a mobile CPU can run them. Pooled energies are accumulated tile by tile, so no sub-band image is ever stored.

The second is the restorer. It is a NAFNet-style U-shaped network at 1–4M parameters. In each block, a single linear layer maps the tile signature to a per-channel scale and shift, bilinearly upsampled to feature resolution. The mapping uses no sigmoid, which keeps the activation-free property. The only learned parameters attributable to conditioning are these small linear maps.

The third is an analysis program. Its job is to establish whether the signature carries the information we claim before any restoration result is trusted. We will train linear probes to predict corruption type and level from the signature, compare it against competing descriptors, and visualize signature geometry for noise, rain, blur and haze across levels the restorer never saw.

All restorers are trained on the standard three-task all-in-one mixture used by AirNet and PromptIR. This is Gaussian denoising at σ ∈ {15, 25, 50} evaluated on BSD68, deraining on Rain100L, and dehazing on RESIDE SOTS. A five-task extension adds GoPro deblurring and LOL-v2 low-light enhancement.

Baselines

Unconditioned backbone. The naive reference point is the identical NAFNet-style backbone trained on the same task mixture with no conditioning. Any gain from our method must exceed this model at matched parameter count. It also shows how much a small network already infers implicitly, and a negligible gap would mean conditioning is unnecessary at this scale.

Learned prompt estimator on the same backbone. The strongest competitor on our actual question is PromptIR's prompt generation and interaction modules, transplanted onto our backbone and given an equal or larger parameter budget. It is the strongest because it is the best-performing learned estimator in the file. The file reports its relative gains (0.86 dB average over AirNet, a 2.64 dB dehazing gain over the previous best, and roughly a 7 dB lead over AirNet at σ = 100 after training at σ = 50). However, no absolute per-task PSNR for PromptIR is reported on Lacuna. The only absolute all-in-one figure in the file is AirNet's 34.90 dB on Rain100L, which we will reproduce as a sanity anchor. Both conditioned models share the data, schedule, augmentation and hyperparameter-search budget. Beating it means matching its in-distribution average within 0.1 dB while using fewer conditioning parameters and lower latency, and exceeding it on the out-of-distribution suites described below.

Full-size published systems. We will also run the published PromptIR and AirNet systems at full size. They serve as upper references for quality and lower references for efficiency. We will quantize PromptIR to 8 bits to give a "compress the big model" baseline.

Descriptor ablations. The isolating ablations replace only the signature and keep everything else fixed. One uses the energies of isotropic wavelet sub-bands (LL, LH, HL, HH), as in PW-FNet's frontend. The other, decisive for the parabolic claim, uses an isotropic multi-orientation bank: Gabor-like filters with the same number of scales, orientations and filters, but a fixed aspect ratio at every scale. A third control feeds a shuffled signature taken from a different image, to verify the network actually uses the conditioning.

Why not just…?

Why not just let a learned prompt module figure out the degradation? This is a real threat for the training degradations, where a learned estimator can fit the three tasks well. Our argument is about what happens outside them. A globally pooled learned embedding has no built-in reason to place σ = 100 noise near σ = 50 noise. Our signature changes smoothly with noise level by construction, because fine-scale energy rises monotonically. The matched-backbone prompt baseline settles this. If it equals our model out of distribution as well, the remaining advantage is cost alone. We will then report the result as an efficiency result.

Why not just use wavelet energies, or an oriented-gradient histogram, which are cheaper still? This is the most dangerous alternative. A multi-scale bank with eight orientations and fixed aspect ratio may capture nearly everything that matters, and the parabolic claim then reduces to "orientation helps." The matched-filter-count isotropic bank tests exactly this. We predict the gap appears mainly on thin rain streaks and on separating fine-scale noise from fine texture, where elongated atoms concentrate streak energy into fewer cells. If no gap appears, we will say so and drop parabolic scaling from the claim. The three-orientation wavelet descriptor should lose more clearly, because it cannot resolve diagonal streak angles at all, consistent with the near-1.0 wavelet sparsity LDAN measured.

Why not just compress PromptIR for the phone? Compression keeps the Restormer attention and the learned estimator, so it inherits their cost. Evidence from Toward DNN of LUTs suggests restoration quality is sensitive to precision. The file does not report a quantized PromptIR, so we will measure one. Its PSNR and its latency on the same mobile CPU directly determine whether a purpose-built small model is needed.

Why not just a classifier that picks among per-task specialist models? This uses hard routing and stores several models. It cannot express mixtures or intermediate levels. We include an oracle version, given the true task label, as a ceiling for in-distribution quality.

Proposed Experiments

The first decisive check precedes restoration. Linear probes predict corruption type and level from each descriptor: the parabolic signature, the isotropic multi-orientation bank, wavelet energies, and the frozen learned prompt embedding. Probes are trained on the training levels and tested on held-out ones (σ = 75 and 100, Rain200H for heavy rain). If the parabolic signature does not beat the isotropic bank at regressing the held-out levels, the parabolic hypothesis fails early and cheaply.

Restoration is then evaluated in two regimes. In distribution, we report PSNR and SSIM on BSD68, Rain100L and SOTS. Out of distribution, we use noise at σ = 100, Rain200H, real rain from SPA-Data and GT-Rain, real non-uniform haze from NH-HAZE, real sensor noise from SIDD without retraining, and synthetic mixtures such as noise plus rain that are never seen in training. The mixture test is where spatially tiled, directional conditioning should help most. The shuffled-signature control should cost several dB; if it does not, the network is ignoring the conditioning.

Efficiency is measured on hardware, not just counted. We will report parameters, multiply-accumulates, peak activation memory, and wall-clock latency on a Snapdragon 888 CPU, the device used by MuLUT and ShiftLUT, for 1024×1024 inputs. The signature extractor's latency is reported separately. Our target is under 5% of the total.

The success criteria are three. The parabolic signature must beat the isotropic bank on held-out-level probes and on the rain and mixture suites. The conditioned small model must match the transplanted prompt baseline in distribution and beat it out of distribution. The full model must run at least several times faster than quantized PromptIR on the same phone CPU.