Telling Mislabeled Radar Chips from Hard Viewing Angles

When a training set has only ten or twenty radar images per vehicle class, one wrong label is a large share of a class's supervision. Methods for learning with noisy labels usually find bad labels by looking for training examples the network finds hard. In synthetic aperture radar (SAR), a microwave imaging system whose pictures of a vehicle change sharply with the vehicle's heading, many correctly labeled examples are also hard: a tank seen end-on scatters very differently from the same tank seen broadside. This proposal argues that small-sample SAR target recognition on the MSTAR benchmark needs a label-cleaning rule that separates "hard because of the viewing angle" from "hard because the label is wrong." MSTAR records the heading angle (azimuth) of every chip, which makes that separation possible.

The method has two parts. First, each chip is scored by how well models that never trained on it predict its given label. Second, that score is corrected for how ambiguous chips at that azimuth are in general, and the ambiguity is measured without using labels. Chips that still score low are not thrown away. They move into the unlabeled pool of a SAR-appropriate semi-supervised learner, so their images keep contributing information about rare viewing angles. The work also builds an evaluation protocol for MSTAR that includes realistic noise types, namely azimuth-concentrated confusions and pseudo-labels from simulated-to-real transfer. Right now no noisy-label method has been tested on SAR data at all.

Core Claim

The obvious way to combine these directions is to take a noise-robust method from optical vision and run it on a small MSTAR subset. The research file finds no such evaluation. Every noisy-label method it covers, including VolMinNet, Patrini et al., D&C-Net, SELC, WDN and Shu et al., was evaluated on CIFAR, Clothing1M and similar data. The SAR scarcity work, including Φ-GAN, Multi-Block Mixed semi-supervised learning (MBM) and AADR, assumes its few labels are correct. The "noise" studied in SAR work such as LPPCL and DRRNet-SKD is sensor speckle, not wrong class labels.

The claim here is narrower and testable. In the 10–40 labels-per-class regime on MSTAR, we expect existing loss-based clean/noisy separation to fail for two reasons. One is that it is statistically unstable at these sample sizes. The other, more interesting reason is that it mistakes clean chips at atypical azimuths for mislabeled ones. A selection score that is computed from held-out predictions and normalized against label-free azimuth difficulty should recover most of the lost accuracy. The largest gains should appear under the two noise types that matter operationally: confusions concentrated at particular aspect angles, and pseudo-labels inherited from a simulation-trained model.

Why the obvious tools break here

Two lines of evidence motivate the proposal. First, the selection-based methods D&C-Net and SELC fit a two-component Gaussian mixture to per-sample training loss. SELC also uses the gap between the two mixture means to decide when memorization begins. Both work with tens of thousands of examples. With 100–400 training chips in total, the loss histogram has no reliable two-peak shape. In addition, the training loss of any chip the network has already fit says little about whether its label is correct.

Second, Drory, Avidan and Giryes show that networks resist label noise because neighboring clean examples outvote the noisy ones in feature space. That protection vanishes when the noise is locally concentrated, and accuracy then falls linearly with the noise rate. Systematic confusion between similar tracked vehicles at broadside is exactly this kind of locally concentrated noise. Pseudo-labels produced by a model trained on simulated data, as in AADR, will also err non-uniformly, most likely at the aspects where simulation and measurement differ most.

The constraints come from the SAR scarcity literature. MBM shows that standard Mixup and CutMix create non-physical radar artifacts, while block-wise mixing (4×4 blocks) reaches 99.56% accuracy with 20 labels per class. This tells us which semi-supervised learner to build on. It also tells us not to borrow the Mixup step that gives SELC's SELC+ variant its strength. Φ-GAN shows that physics constraints are valuable for generation under scarcity. We deliberately do not put a scattering-physics model into the label-cleaning step, for reasons given below.

Research Program

The method has three stages, all running on the labeled set $\mathcal{D}=\{(x_i, y_i, \theta_i)\}$ of chips $x_i$, possibly wrong labels $y_i$ and recorded azimuths $\theta_i$.

The first stage is cross-fitting. The labeled set is split into $K$ folds, stratified by both class and azimuth sector. A small classifier is trained on $K-1$ folds and predicts the held-out fold, and this is repeated over a few seeds. Each chip then has a held-out class distribution $p_{-i}(\cdot \mid x_i)$ from models that never saw its label. At MSTAR chip sizes and these sample counts, ten-fold cross-fitting costs about as much as a single ordinary training run on the full dataset. Using held-out predictions directly answers the memorization problem that undermines train-loss heuristics at small $n$.

The second stage is azimuth-difficulty normalization. For each azimuth window, we compute the mean entropy $\bar{H}(\theta)$ of held-out predictions, pooled over every chip in that window regardless of its label. Each MSTAR class is imaged around the full circle, so every window contains chips from all classes. The selection score is

$$s_i = \log p_{-i}(y_i \mid x_i) - g\big(\bar{H}(\theta_i)\big),$$

where $g$ is a monotone function fitted by robust regression of label agreement on window entropy across the whole set. Here $s_i$ measures how much worse the given label is predicted than a typical chip at the same aspect.

The difficulty term is label-free on purpose. If difficulty were measured by agreement with labels inside each azimuth window, a cluster of concentrated mislabels would raise that window's apparent difficulty and normalize itself away. Entropy instead rises when the images at an aspect are ambiguous, not when the labels are wrong. Whether this separation holds in practice is one of the decisive checks below.

The third stage is demotion. Chips whose $s_i$ falls below a threshold are not discarded. Their labels are removed and they join the unlabeled pool of an MBM-style semi-supervised learner, where the learner's own pseudo-labeling and block mixing use their images. The threshold is chosen from the clean/noisy separation of $s$ itself. A two-component mixture is fitted to $s$ pooled over all classes, not per class, which makes the fit far more stable than per-class loss fitting. A fixed-quantile fallback is reported alongside it. Moving demoted chips to the unlabeled pool rather than deleting them addresses the concern that discarding hard examples starves the model of viewing angles.

Baselines

The naive reference point is MBM trained on all noisy labels with no cleaning. It shows how much damage each noise type does to the best SAR-specific small-sample learner. The upper reference is MBM trained on only the truly clean subset, with mislabeled chips moved to the unlabeled pool by an oracle, which bounds what any selection rule can achieve. The best existing clean-label number we can anchor to is MBM's reported 99.56% accuracy at 20 labels per class (MBM). The oracle reproduction should land near that figure. If it does not, our reimplementation is suspect.

The strongest existing noisy-label competitors are SELC and D&C-Net. SELC corrects labels with a running average of the model's own predictions and reports 78.62% on CIFAR-10 with 80% symmetric noise. D&C-Net splits clean from noisy chips by loss and corrects only the noisy ones, and reports 87.30% on CIFAR-10 with 50% symmetric noise. We include VolMinNet, which estimates a class-to-class noise matrix without anchor points and reports 88.54% on CIFAR-10 with 45% pair-flip noise, because pair-flip confusion between similar vehicles is its home ground. We found no reported result on Lacuna for any of these methods on MSTAR or any SAR dataset, so all MSTAR numbers will be produced here.

Fairness requires identical backbones, identical data splits and noise draws, and the MBM semi-supervised learner wrapped around every method wherever it applies. Each method also gets the same budget of 20 hyperparameter configurations, selected on a held-out noisy validation split, never a clean one. The proposal beats these methods if it improves 15°-depression test accuracy by at least three points over the best of them, averaged across seeds, under azimuth-concentrated and pseudo-label noise at 10 and 20 labels per class, without losing more than one point under symmetric noise.

Four ablations isolate the key ingredient. The first uses cross-fitting without azimuth normalization, which is essentially confident-learning-style selection and is the most important comparison. The second applies azimuth normalization to training loss instead of held-out predictions. The third computes difficulty from label agreement rather than label-free entropy. The fourth deletes demoted chips rather than moving them to the unlabeled pool.

Why not just…?

Why not just run SELC or D&C-Net on MSTAR? This is the most direct threat, and the baseline comparison settles it. Our prediction rests on the specific point that their loss mixtures are fit on memorizable training losses at a scale they were never tested at. If SELC's running-average correction proves stable at 20 labels per class and does not reject clean end-on chips, the proposal's contribution shrinks to the benchmark itself. We will report that outcome as such.

Why not just normalize training loss within azimuth bins? This is the simplest way to account for viewing angle, and it is ablation two. It should fail for two reasons. Training loss on tiny sets is near zero for most chips, mislabeled ones included, once fitting completes. And label-dependent normalization hides concentrated noise. If this ablation matches the full method, the cross-fitting stage is unnecessary and we will say so.

Why not just reweight with a small clean validation set, as in Shu et al.? Shu et al. weight each training example by how well its gradient aligns with the gradient on as few as ten clean examples per class. With only 10–20 labels per class, reserving ten verified clean chips per class consumes most of the labeling budget. It also assumes those chips are clean, and that assumption is what fails when labels come from a simulation-trained model. We run it in a privileged variant that is given a verified clean set as an extra resource. This tells us how much our method gives up by not having one.

Why not build the scattering physics into a noise matrix that depends on azimuth? That design appeals because it combines Φ-GAN's scattering-center model with Patrini et al.'s noise matrix. However, a 10×10 matrix per azimuth sector has more free parameters than there are labeled chips. We defer it unless the azimuth-difficulty diagnostic below shows that confusions really are aspect-specific in a way a single global matrix cannot capture.

Proposed Experiments

The data are MSTAR standard operating conditions: ten classes, with chips at 17° depression angle for training and 15° for testing. Labeled budgets are 10, 20 and 40 per class, plus the 10% split used by Φ-GAN. The remaining training chips serve as unlabeled data. A depression-shift test, training at 17° and testing at 30°, checks whether cleaning helps under extended operating conditions.

There are four noise families. Symmetric noise is applied at 20% and 40%. Pair-flip noise is applied at 20% and 40% among visually similar vehicles (BMP2, BTR70, T72). Azimuth-concentrated noise flips those same pairs only within 20-degree windows around broadside or end-on, which mimics an annotator's systematic confusion. The fourth family is pseudo-label noise: labels for the MSTAR chips come from a classifier trained on the simulated SAMPLE dataset, the setting AADR uses without correction. The metrics are test accuracy over five seeds, precision and recall of noisy-chip detection, and recall of clean chips by azimuth sector. The last metric directly measures whether hard aspects survive the cleaning step.

The first decisive check tests the premise. With clean labels at 20 per class, does the held-out confidence on true labels drop systematically with azimuth? If it is flat across sectors, azimuth-induced difficulty is not a real confound on MSTAR and the normalization stage should be dropped. The second check tests whether label-free entropy behaves as intended: under azimuth-concentrated noise it should not rise inside the corrupted windows, even though label agreement does. The third check is the head-to-head comparison against SELC, D&C-Net and VolMinNet under the realistic noise families. The fourth is the ablation pair comparing deletion with demotion, measured by accuracy at the rarest test azimuths.