Radar Intelligent Detection with Coarse-Grained Labels

I. INTRODUCTION

• Statistical detectors can degrade when clutter violates their assumed distribution, while feature-based methods may generalize poorly across clutter conditions and target types.These limitations motivate approaches that adapt to complex radar environments.Source: §I · p. 1
• Deep-learning radar detectors include CNN, graph-neural-network, and LSTM approaches, with prior examples using attention-enhanced, spatiotemporal, and sequence-based models.The introduction also notes that these detectors are typically trained with range-cell-level labels.Source: §I · p. 1
• Range-cell labels are expensive to obtain, whereas range-window labels omit the target’s position within the window.The paper identifies accurate range-cell detection under this coarser supervision as an under-investigated problem in radar detection.Source: §I · p. 1
• The proposed method combines a parameter-sharing siamese Transformer with a shift-equivariant consistency constraint for training on range-window labels.The constraint uses the range-dimension shift prior to guide precise target localization from paired views.Source: §I · p. 1

A. Signal Model

• The transmitted signal is modeled as a linear frequency-modulated waveform, with carrier frequency f_c, chirp rate k, and pulse width T_r.For a point target at range R, the paper then introduces the received signal; τ=2R/c0, with c0 the speed of light.Source: A. Signal Model · p. 1
• After standard signal processing, the baseband echo r has M range cells, and the target is assumed to lie at cell m0=⌈τf_s⌉.Here f_s is the sampling rate.Source: Figure 1 · p. 2
• The model uses only a window label, which indicates target presence but not its location, so it pairs two crops with a known relative shift to constrain consistent target locations.The paired views share the label and are processed by a parameter-sharing siamese Transformer with an SEC constraint.Source: A. Signal Model · p. 2
• A labeled window is extended by L range cells on each side; each window is assumed to contain at most one target, and the added cells are assumed target-free.Two length-K crops include the same length-S labeled window, with S=K−L.Source: A. Signal Model · p. 2
• Each crop stacks its real and imaginary parts as two channels before feeding them to the two Transformer branches.The relative shift is determined by the crops’ offsets and is known during training.Source: A. Signal Model · p. 2

B. Siamese Transformer

• The siamese Transformer has two identical branches with shared parameters, each containing feature embedding, positional encoding, Transformer blocks, and a classifier.A convolutional layer maps the input X to a feature representation F.Source: B. Siamese Transformer · p. 2
• Because Transformer blocks are insensitive to absolute input positions, a learned encoding of the range sequence is added to the embedded features.The positional sequence h=[1,2,…,K] is mapped through a fully connected layer to P, which is added to F.Source: B. Siamese Transformer · p. 2
• Each Transformer block uses pre-normalization, multi-head attention, a feedforward network, and residual connections.Attention integrates information across range cells; the feedforward network enhances target–clutter separability.Source: B. Siamese Transformer · pp. 2–3
• The classifier produces a range-cell-wise score vector and applies a sigmoid to obtain posterior probabilities, where p_k is the probability that cell k contains a target.The two echo views produce separate posterior-probability vectors.Source: B. Siamese Transformer · p. 3

• For a negative window, the pseudo label is all zero; for a positive window, the highest-posterior cell is labeled target and the rest non-target.Source: C. Shift-Equivariant Consistency… · p. 3
• The known relative crop offset shifts view 1’s pseudo label to view 2, encoding the expectation that predicted responses shift with the input.Source: C. Shift-Equivariant Consistency… · p. 3
• The objective equally averages the two views’ per-cell binary cross-entropy losses, encouraging sharp responses and cross-view consistency.The paper says this constraint improves response-pattern reliability and detection performance.Source: C. Shift-Equivariant Consistency… · p. 3

• Only one branch of the siamese Transformer is retained as the detector at inference.Source: D. Inference · p. 3
• The input echo is partitioned along range with a sliding window of length K, and each window is processed sequentially to obtain cell-wise posterior probabilities.Source: D. Inference · p. 3
• A threshold h = 0.5 is applied to each cell’s posterior probability to produce the binary detection vector.Source: D. Inference · p. 3

• Point targets were simulated using radar parameters and superimposed on real sea clutter to control target strength while retaining measured clutter characteristics.Clutter-only backgrounds came from range cells 130–450 and 680–1000; targets were placed at cells 300 and 850.Source: A. Experimental Setup · p. 3
• The dataset was divided into training, validation, and test sets at a 6:2:2 ratio.Source: A. Experimental Setup · p. 3
• A detection counted as correct if its predicted range cell matched ground truth or fell within one adjacent cell on either side.Source: A. Experimental Setup · p. 3

• The component ablations were conducted at SCR = 10 dB and summarized in Table II.Source: B. Effectiveness Evaluation · p. 3, Table II · p. 4
• Removing either the siamese architecture or SEC prevents convergence to an effective detector; retaining both with CV-UNet yields comparable detection performance.The ablation results are summarized in Table II.Source: Table II · p. 4, B. Effectiveness Evaluation · p. 4
• The “✗Siamese” variant trains only view 1 with gradients updating shared parameters, while “✗Transformer” substitutes CV-UNet for the Transformer backbone.Source: Figure 2 · p. 4

C. Performance Comparison

• At SCR=5 dB, the range-cell-labeled reference uses the same network architecture and inference pipeline as the proposed detector, differing only in annotation detail.This makes it a performance boundary for assessing the effect of replacing fine-grained labels with range-window labels.Source: C. Performance Comparison · p. 4
• Across 0–11 dB SCR, the proposed method exceeds all four CFAR variants while keeping false alarms low, but matches the cell-label reference only from 3 dB onward.A performance gap remains at 0–2 dB SCR.Source: C. Performance Comparison · p. 4, Figure 3 · p. 4
• Using range-window labels achieves performance close to range-cell-label training in most SCR conditions, substantially reducing fine-grained labeling reliance and data-preparation effort.The authors report little loss in detection performance and easier practical deployment.Source: C. Performance Comparison · p. 4

IV. CONCLUSION

• The proposed radar target detection training method requires only coarse range-window labels, addressing the practical difficulty of obtaining fine-grained range-cell labels.The paper identifies reduced annotation difficulty and data-preparation cost as benefits.Source: §IV · p. 4
• A siamese Transformer learns paired views from echoes in the same range window, while a shift-equivariant consistency constraint incorporates the range-dimension shift prior.The approach aims to learn range-cell-level target detection responses under window-level supervision.Source: §IV · p. 4
• The method significantly outperforms classical CFAR detectors and achieves performance comparable to fully supervised methods with fine-grained labels under most SCR conditions.The conclusion characterizes this as little sacrifice in detection performance alongside reduced labeling cost.Source: §IV · p. 4