Wrong labels bias the supervisory signal
Deep-learning radar target detection has substantially improved detection performance, but the paper states that its effectiveness depends critically on the correctness of the training labels assigned to radar echo samples. In practical applications, training data may contain partially mislabeled samples, which can cause detection models to learn biased supervisory information and thereby degrade detection performance.
The paper frames this as a label-noise problem specific to radar target detection rather than a generic classification issue. It establishes a probabilistic model to characterize the generation mechanisms of label errors in radar target detection, then analyzes the robustness of detection models under partial label errors within a risk minimization formulation.
A symmetry condition makes the loss robust
The theoretical core of the paper is a condition on the loss function. The analysis shows that, under appropriate assumptions, a loss function satisfying the symmetry condition enables the detection model in the learning framework to be robust against partial label errors.
Guided by this result, the authors design a loss function satisfying the symmetry condition and construct the proposed robust learning framework, SL-RLF. Theoretically, under the corresponding label-error conditions, the optimal detection model learned by SL-RLF from partially mislabeled training data can achieve the same detection performance under the true data distribution as the optimal model learned from accurately labeled data.
This is a statement about optimal models under stated conditions, not a guarantee that any trained network will reach that optimum. The paper treats the symmetry condition as a sufficient condition for robustness, which matters when reading the experimental thresholds below.
Three label-error types set different thresholds
The experiments separate label errors into distinct mechanisms rather than one noise level. Type 1 is a constant uniform error model, where labels are perturbed with a fixed probability. Type 2 is a sample-dependent uniform error, where each sample's error probability is produced by a function g(x) whose output lies in [0,1] and whose upper bound is controlled by a scaling factor alpha. Type 3 is a class-dependent error, where the clutter class and target class have their own error probabilities.
These types carry different robustness ranges. For Type 1, the admissible range of the error probability phi_0 expands as the number of classes K increases, consistent with the condition phi_0 < (K-1)/K. For Type 2, the required condition is max over samples of phi_xi = alpha < (K-1)/K, which equals 0.5 in the two-class setting. For Type 3 in a two-class detection task, the effective robustness range simplifies to phi_yxi < 0.5.
Because the thresholds differ by mechanism, a single reported error rate cannot be compared across types without naming which mechanism produced it.
SL-RLF outperforms Co-teaching and TCE
Across the tested settings, the paper reports that SL-RLF consistently outperforms representative methods, including Co-teaching and TCE, under different types and levels of partial label errors, and that its robustness is generally consistent with the theoretical analysis.
Under sample-dependent uniform errors, when the condition alpha < 0.5 required by the corresponding theorem is satisfied, SL-RLF exhibits strong robustness and achieves overall better performance than Co-teaching and TCE. Under class-dependent errors in the two-class task, SL-RLF still maintains superior robustness compared with those methods.
The comparison is against these named baselines on the paper's radar datasets and error constructions, so the ranking is specific to those conditions rather than a general claim about all noisy-label methods.
Performance degrades near the theoretical boundaries
The paper is explicit that its theorems give sufficient conditions rather than necessary ones. When alpha is at or above 0.5 under sample-dependent uniform errors, SL-RLF still retains a certain degree of robustness; the theoretical guarantee no longer holds, but robustness does not immediately fail.
The behavior near the boundary differs by error type. Under sample-dependent uniform errors, SL-RLF does not show an abrupt performance change near alpha = 0.5, because alpha only controls the upper bound of each sample's error probability, and some samples still satisfy phi_xi < 0.5 and continue to provide effective supervisory information. Under constant uniform errors, once phi_0 = 0.5, the labels of all samples are flipped with probability 0.5, so the training labels no longer contain effective discriminative information about the true labels.
Under class-dependent errors, when either the clutter-class error probability phi_c or the target-class error probability phi_t approaches 0.5, performance degrades pronouncedly. The paper attributes the asymmetry between the two classes to class imbalance: because clutter samples greatly outnumber target samples, flipping target labels at 0.5 affects mainly the smaller class, leaving many correctly labeled clutter samples, whereas flipping clutter labels at 0.5 degrades the dominant source of supervision.
Robustness holds across classes, functions, and architectures
Several ablations test whether the framework's robustness depends on incidental choices. For the number of target classes, SL-RLF maintains robustness under different class-number settings, and the admissible error-probability range expands as the number of classes increases, consistent with the theoretical condition.
For the form of the sample-dependent function g, the paper runs ablations with four distributions whose outputs follow uniform, logarithmic, Beta, and truncated Gaussian distributions. Within the range alpha < 0.5 required by the corresponding theorem, SL-RLF maintains stable robustness for different types of g, confirming that the guarantee does not depend on the specific form of g. When alpha > 0.5, the degree of performance degradation differs across choices of g, because different g induce different distributions of error probability and thus different proportions of samples with phi_xi > 0.5.
For model architecture, the paper builds VGG- and ResNet-based detectors and trains them with SL-RLF on sea state 5, HH polarization data with constant uniform errors. Across different network architectures and model depths, training with SL-RLF improves robustness against partial label errors.
Higher sea states weaken robustness
Using data across different polarization modes and sea-state levels, the paper reports that SL-RLF maintains favorable robustness on data corresponding to different polarization modes and sea-state conditions, and that the observed error-probability ranges are consistent with the theoretical range.
The trend across sea states is not flat. As the sea state increases, the robustness of SL-RLF shows a decreasing trend. The paper explains this by noting that the sea-clutter background becomes more complex under higher sea states and target echoes are more likely to be submerged in strong sea clutter, reducing target-background separability and making effective decision boundaries harder to learn. In that setting, the adverse effect of partial label errors on model parameter learning is further amplified.
This explanation is the authors' interpretation of the observed trend, and it identifies sea state as a condition that shifts how much label error the framework can absorb.
Match the error mechanism before trusting the guarantee
The most decision-relevant distinction in this paper is not a single accuracy number but which label-error mechanism is present and how close its error probability sits to the relevant threshold. Constant uniform errors have a sharp boundary at phi_0 = 0.5 in the two-class case, where all labels are flipped with probability 0.5 and effective discriminative information disappears. Sample-dependent uniform errors are governed by alpha < 0.5, but because alpha only bounds each sample's error probability, some samples can remain below 0.5 and keep supplying useful supervision, so degradation is gradual rather than abrupt. Class-dependent errors are governed by phi_yxi < 0.5, and the paper shows the clutter-class error rate matters more than the target-class rate under class imbalance.
A reader deciding whether to rely on SL-RLF should therefore first identify the error type, then check the error probability against the matching threshold, and then account for sea state, since higher sea states reduce robustness. The paper's own framing supports this order: the theorems provide sufficient conditions, the experiments confirm the ranges for each type, and the reported advantage over Co-teaching and TCE holds within those tested radar scenarios rather than as an unconditional claim.
Sources used in this answer
A Robust Learning Framework for Deep-Learning-Based Radar Target Detection With Partially Mislabeled Training Data
SL-RLF, built on a loss function satisfying a symmetry condition, keeps deep-learning radar target detection robust under partial label errors and consistently outperforms Co-teaching and TCE across the tested error types and levels.