Collaborative Scattering–Deep Features (CLSDF)

• The paper introduces a hybrid framework that merges physics-aware scattering representations with deep learning to enhance SAR target recognition under noisy labels.
• CLSDF leverages structured ASC-based scattering features and graph convolution techniques alongside deep image features to extract invariant and discriminative representations.
• Empirical evaluations on the MSTAR dataset demonstrate CLSDF’s superior robustness to various noise levels compared to traditional methods.

Collaborative learning of scattering and deep features (CLSDF) denotes a hybrid representation-learning paradigm in which scattering-derived descriptors and learned deep features are combined within a single trainable system. In its explicit named form, CLSDF is a synthetic aperture radar (SAR) automatic target recognition framework for noisy labels that fuses deep image features from amplitude images with scattering features derived from attributed scattering centers (ASCs), and couples that fusion to class-wise noise modeling, semi-supervised peer training, and joint distribution alignment (Fu et al., 11 Aug 2025). In a broader literature sense, the term also captures a line of work in which scattering representations are not treated as isolated fixed front ends, but are fused, adapted, or jointly optimized with deep modules.

1. Scattering representations as the substrate

The conceptual substrate of CLSDF is the scattering transform. Early wavelet-scattering networks replaced learned early convolutional layers with analytically defined operators built from wavelet transforms, complex modulus nonlinearities, and low-pass averaging. A canonical two-layer formulation concatenates zeroth-, first-, and second-order coefficients as

$Sx=\{S_0x,S_1x,S_2x\},$

with the architecture explicitly framed as a possible replacement or initialization for the first layers of deep networks (Oyallon et al., 2013).

A more expressive scattering construction was later developed for object classification by introducing predefined wavelet filters over both spatial and angular variables. In that setting, the second-order roto-translation scattering representation is written as

$S_J x = A_J\, |W_2|\, |W_1|\, x,$

and is designed to be locally invariant to translations, stable to small deformations, and sensitive to rotation variability without imposing full rotation invariance (Oyallon et al., 2014). The same line of work states that first-order scattering is “very similar to SIFT,” while second-order roto-translation coefficients encode interactions between scales and angles in multiscale neighborhoods (Oyallon et al., 2014).

These properties explain why scattering remains attractive in hybrid systems. It contributes mathematically structured geometric priors, contractivity, and deformation stability, while deep networks contribute task-adaptive abstraction. CLSDF inherits precisely this division of labor: scattering encodes structured invariants that are expensive or unstable to learn from scratch, whereas deep modules supply discriminative adaptation.

2. From fixed front ends to hybrid scattering–deep systems

The transition from fixed scattering pipelines to hybrid scattering–deep architectures proceeded through several distinct strands. Visualization work on ScatterNets showed that higher-order coefficients are often sensitive to complex, edge-like patterns such as checkerboards and rippled edges, and that these responses are “very dissimilar” to second- and third-layer CNN features (Cotter et al., 2017). This diagnosis is important because it frames scattering and deep features as complementary rather than redundant.

A more integrated hybridization appeared in the learnable ScatterNet literature. There, each scattering order is decomposed into a locally invariant layer consisting of fixed wavelet analysis, modulus or low-pass propagation, and a learned cross-channel mixing operator,

$Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$

with learned layers permitted before the scattering stack, between scattering orders, and after the scattering stack, all trained end to end (Cotter et al., 2019). The wavelet filters remain fixed, but the representation becomes learnable through the mixing matrix $A$ and surrounding convolutional layers.

A different hybridization strategy kept the scattering front end fixed and learned only the later representation. In sparse scattering classification, scattering coefficients are first computed and then processed by a learned dictionary and classifier, with sparse coding implemented by an unrolled homotopy thresholding network (Zarka et al., 2019). This preserves scattering’s invariance-inducing structure while shifting learning capacity to a sparse discriminative back end.

An additional, weaker precursor arose in optical inverse problems. A single U-net was trained on a blended dataset of speckle–reference pairs from two distinct scattering media, namely a glass diffuser and a multimode fiber, so that one shared network reconstructed both domains (Yang et al., 2018). The same source explicitly notes that this is better characterized as multi-medium joint supervised reconstruction than as a full collaborative-feature-learning framework (Yang et al., 2018).

Across these works, “collaboration” usually meant either fixed scattering plus a learned back end, or learned adapters around fixed scattering operators. What remained uncommon was direct end-to-end learning of the scattering filters themselves.

3. Formal structure of CLSDF for SAR target recognition

The explicit CLSDF framework addresses SAR automatic target recognition under noisy labels. The training set is written as

$$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$

where $x_n$ is complex SAR data and $y_n$ is its class label (Fu et al., 11 Aug 2025). Each SAR sample is converted into two modalities: an amplitude image $x_I\in\mathbb{R}^{H\times W}$ and an ASC set $x_S\in\mathbb{R}^{P\times 7}$, with $P=40$ selected through parameter analysis (Fu et al., 11 Aug 2025).

In this framework, “scattering” refers not to wavelet-scattering coefficients but to a physical scattering-center representation. The target’s total scattered field is modeled as

$S_J x = A_J\, |W_2|\, |W_1|\, x,$0

with

$S_J x = A_J\, |W_2|\, |W_1|\, x,$1

where each scattering center has parameter vector

$S_J x = A_J\, |W_2|\, |W_1|\, x,$2

These attributes encode amplitude, position, geometry dependence, distributed length, direction, and aspect dependence (Fu et al., 11 Aug 2025).

The ASC set is treated as dynamic graph-structured data. A local graph is initialized by connecting each scattering center to its $S_J x = A_J\, |W_2|\, |W_1|\, x,$3 nearest neighbors, and the edge feature between nodes $S_J x = A_J\, |W_2|\, |W_1|\, x,$4 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$5 is defined as

$S_J x = A_J\, |W_2|\, |W_1|\, x,$6

with the first graph-layer update

$S_J x = A_J\, |W_2|\, |W_1|\, x,$7

The graph is dynamic because later $S_J x = A_J\, |W_2|\, |W_1|\, x,$8-nearest-neighbor relations are recomputed in feature space rather than kept fixed (Fu et al., 11 Aug 2025). If there are $S_J x = A_J\, |W_2|\, |W_1|\, x,$9 graph layers, their outputs are concatenated and globally pooled:

$Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$0

The deep branch operates on the amplitude image. With ResNet-18 as backbone, the deep image feature is written generically as

$Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$1

During training, amplitude images are cropped to $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$2, with random $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$3 crops taken inside a centered $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$4 region and random changes in brightness, contrast, and saturation (Fu et al., 11 Aug 2025).

Feature fusion is strictly feature-level concatenation:

$Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$5

Each branch of CLSDF contains a ResNet-18 deep image feature extractor, a DGCNN with 3 EdgeConv layers for ASC processing, a concatenation-based fusion module, and a classifier $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$6 (Fu et al., 11 Aug 2025). The framework uses two divergent branches and ensembles them at test time.

4. Collaborative learning under noisy labels

The collaborative component of CLSDF has two levels. The first is feature collaboration, namely the fusion of $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$7 and $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$8. The second is branch collaboration, in which two branches exchange their estimates of which samples are clean or noisy and use those peer partitions for semi-supervised training (Fu et al., 11 Aug 2025).

Sample selection is class-conditional. For class $Y^{(l+1)}(u)=A\,Z^{(l+1)}(u),$9, the per-sample losses are collected as

$A$0

Instead of fitting one global loss model, CLSDF fits a separate two-component Gaussian mixture model for each class. The posterior probability that a sample is clean is denoted $A$1, using the Gaussian component with smaller mean. Partitioning then follows

$A$2

The threshold is $A$3 (Fu et al., 11 Aug 2025).

On the clean subset, CLSDF performs label co-refinement:

$A$4

where $A$5 augmentations are used. On the noisy subset, it performs label co-guessing with both branches:

$A$6

The guessed label is then calibrated by joint distribution alignment,

$A$7

so that the joint distribution of clean labels and guessed noisy labels better matches the target marginal class distribution (Fu et al., 11 Aug 2025).

Both refined and guessed labels are sharpened with temperature $A$8. The clean and noisy subsets are then mixed by a MixMatch-style interpolation,

$A$9

$$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$0

with $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$1. Training uses cross-entropy on the mixed clean subset and mean-squared error on the mixed noisy subset:

$$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$2

$$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$3

and the total loss is

$$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$4

The unlabeled loss weight $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$5 is linearly ramped up to $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$6 over the first $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$7 epochs (Fu et al., 11 Aug 2025).

Optimization uses SGD with learning rate $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$8, batch size $$\mathcal{D}=(\mathcal{X},\mathcal{Y})=\{(x_n,y_n)\}_{n=1}^N,$$9, total training length $x_n$0 epochs, and a $x_n$1-epoch warm-up before collaborative semi-supervised learning begins. The implementation is in PyTorch on Nvidia RTX 3090 hardware (Fu et al., 11 Aug 2025).

5. Empirical behavior, robustness, and ablations

The principal empirical evaluation is on the MSTAR dataset, collected by the Sandia National Laboratory SAR sensor platform, with X-band, HH polarization, and 10 target classes: BMP2, BTR70, T72, BTR60, 2S1, BRDM2, D7, T62, ZIL131, and ZSU234 (Fu et al., 11 Aug 2025). Under standard operating condition (SOC), training uses depression angle $x_n$2 and testing uses $x_n$3. Extended operating conditions include large depression changes, target-version variation, and additive white Gaussian noise at SNRs $x_n$4 dB (Fu et al., 11 Aug 2025).

Under SOC with symmetric label noise, CLSDF reports $x_n$5, $x_n$6, $x_n$7, and $x_n$8 at noise rates $x_n$9, $y_n$0, $y_n$1, and $y_n$2, respectively. In the same setting, ELR reports $y_n$3, $y_n$4, $y_n$5, and $y_n$6, while DivideMix reports $y_n$7, $y_n$8, $y_n$9, and $x_I\in\mathbb{R}^{H\times W}$0 (Fu et al., 11 Aug 2025). Under asymmetric noise, CLSDF reports $x_I\in\mathbb{R}^{H\times W}$1, $x_I\in\mathbb{R}^{H\times W}$2, $x_I\in\mathbb{R}^{H\times W}$3, and $x_I\in\mathbb{R}^{H\times W}$4 at $x_I\in\mathbb{R}^{H\times W}$5, $x_I\in\mathbb{R}^{H\times W}$6, $x_I\in\mathbb{R}^{H\times W}$7, and $x_I\in\mathbb{R}^{H\times W}$8, compared with $x_I\in\mathbb{R}^{H\times W}$9, $x_S\in\mathbb{R}^{P\times 7}$0, $x_S\in\mathbb{R}^{P\times 7}$1, and $x_S\in\mathbb{R}^{P\times 7}$2 for DivideMix (ASC+Image), and $x_S\in\mathbb{R}^{P\times 7}$3, $x_S\in\mathbb{R}^{P\times 7}$4, $x_S\in\mathbb{R}^{P\times 7}$5, and $x_S\in\mathbb{R}^{P\times 7}$6 for Co-learning (Fu et al., 11 Aug 2025).

The same robustness extends to more difficult operating conditions. Under EOC-2, CLSDF reports $x_S\in\mathbb{R}^{P\times 7}$7, $x_S\in\mathbb{R}^{P\times 7}$8, $x_S\in\mathbb{R}^{P\times 7}$9, and $P=40$0 at $P=40$1, $P=40$2, $P=40$3, and $P=40$4 symmetric noise, whereas Cross-Entropy reports $P=40$5, $P=40$6, $P=40$7, and $P=40$8, DivideMix reports $P=40$9, $S_J x = A_J\, |W_2|\, |W_1|\, x,$00, $S_J x = A_J\, |W_2|\, |W_1|\, x,$01, and $S_J x = A_J\, |W_2|\, |W_1|\, x,$02, and ELR reports $S_J x = A_J\, |W_2|\, |W_1|\, x,$03, $S_J x = A_J\, |W_2|\, |W_1|\, x,$04, $S_J x = A_J\, |W_2|\, |W_1|\, x,$05, and $S_J x = A_J\, |W_2|\, |W_1|\, x,$06 (Fu et al., 11 Aug 2025). For EOC-1 and EOC-4, the reported evidence is plot-based rather than tabulated in the supplied text, but the stated trend is that performance declines as angle difference or test-image corruption increases, with CLSDF remaining superior (Fu et al., 11 Aug 2025). On EOC-3, confusion matrices show reduced confusion across target versions relative to Cross-Entropy and DivideMix. The method is also validated on SAR-ACD, where it is reported to achieve advanced performance under $S_J x = A_J\, |W_2|\, |W_1|\, x,$07 symmetric label noise (Fu et al., 11 Aug 2025).

Ablation studies separate the method into multi-model feature extraction (MMFE), class-wise sample selection (CWSS), and joint distribution alignment (JDA). At $S_J x = A_J\, |W_2|\, |W_1|\, x,$08 noise under SOC, DivideMix yields $S_J x = A_J\, |W_2|\, |W_1|\, x,$09 symmetric and $S_J x = A_J\, |W_2|\, |W_1|\, x,$10 asymmetric accuracy. Adding the CLSDF training framework without CWSS and JDA gives $S_J x = A_J\, |W_2|\, |W_1|\, x,$11 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$12; removing only JDA gives $S_J x = A_J\, |W_2|\, |W_1|\, x,$13 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$14; removing only CWSS gives $S_J x = A_J\, |W_2|\, |W_1|\, x,$15 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$16; removing MMFE gives $S_J x = A_J\, |W_2|\, |W_1|\, x,$17 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$18; and full CLSDF reaches $S_J x = A_J\, |W_2|\, |W_1|\, x,$19 and $S_J x = A_J\, |W_2|\, |W_1|\, x,$20 (Fu et al., 11 Aug 2025). Parameter studies further select $S_J x = A_J\, |W_2|\, |W_1|\, x,$21, $S_J x = A_J\, |W_2|\, |W_1|\, x,$22, and $S_J x = A_J\, |W_2|\, |W_1|\, x,$23 EdgeConv layers as the preferred tradeoffs (Fu et al., 11 Aug 2025).

6. Scope, interpretation, and limitations

CLSDF should be understood as a specific kind of hybrid scattering–deep system. In the named SAR framework, collaboration occurs through two mechanisms: feature collaboration, via concatenation of ASC-based scattering features and deep image features, and branch collaboration, via peer-provided clean/noisy partitions, co-refinement, co-guessing, and aligned pseudo-labeling (Fu et al., 11 Aug 2025). The scattering component is therefore physics-aware and graph-structured, rather than a conventional wavelet-scattering tensor.

This distinguishes CLSDF from several adjacent notions in the earlier literature. It is not merely a fixed scattering front end followed by a learned classifier, as in fixed-scattering pipelines. It is also not a learnable wavelet-scattering architecture in which learned mixing is inserted between scattering orders. Instead, its scattering side is the ASC representation itself, and its deep side is the amplitude-image branch. The result is a multi-model fusion system embedded inside a collaborative noisy-label-learning procedure.

Several limitations are explicit or implicit in the formulation. ASC extraction must be available and meaningful; the fixed number of scattering centers $S_J x = A_J\, |W_2|\, |W_1|\, x,$24 may either include clutter when the true number of scattering centers is smaller or miss useful information when it is larger; the two-branch CNN+DGCNN pipeline, repeated augmentations, and per-epoch GMM fitting introduce computational overhead; and the reported experiments address closed-set label noise rather than open-set noisy labels (Fu et al., 11 Aug 2025). The method also assumes that class-wise loss distributions are separable enough for two-component GMM fitting and that marginal class-distribution estimates used in joint distribution alignment remain informative.

Within the broader history of scattering–deep hybrids, CLSDF is best regarded as a physics-aware collaborative fusion framework rather than as a method for learning scattering filters. A plausible implication is that future work under the same conceptual umbrella may differ mainly in where the scattering prior enters—fixed wavelet modules, physical scattering-center models, or learned scattering adapters—while retaining the same basic premise: structured scattering priors and learned deep features can be more effective together than in isolation.