ZoomSpec: A Physics-Guided Coarse-to-Fine Framework for Wideband Spectrum Sensing

Zhentao Yang; Yixiang Luomei; Zhuoyang Liu; Zhenyu Liu; Feng Xu

cs.CVApr 15, 2026

ZoomSpec: A Physics-Guided Coarse-to-Fine Framework for Wideband Spectrum Sensing

Authors: Zhentao Yang, Yixiang Luomei, Zhuoyang Liu, Zhenyu Liu, Feng Xu

Organizations: Key Laboratory for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China

Abstract

Wideband spectrum sensing for low-altitude monitoring is critical yet challenging due to heterogeneous protocols,large bandwidths, and non-stationary SNR. Existing data-driven approaches treat spectrograms as natural images,suffering from domain mismatch: they neglect time-frequency resolution constraints and spectral leakage, leading topoor narrowband visibility. This paper proposes ZoomSpec, a physics-guided coarse-to-fine framework integrating signal processing priors with deep learning. We introduce a Log-Space STFT (LS-STFT) to overcome the geometric bottleneck of linear spectrograms, sharpening narrowband structures while maintaining constant relative resolution. A lightweight Coarse Proposal Net (CPN) rapidly screens the full band. To bridge coarse detection and fine recognition, we design an Adaptive Heterodyne Low-Pass (AHLP) module that executes center-frequency aligning, bandwidth-matched filtering, and safe decimation, purifying signals of out-of-band interference. A Fine Recognition Net (FRN) fuses purified time-domain I/Q with spectral magnitude via dual-domain attention to jointly refine temporal boundaries and modulation classification. Evaluations on the SpaceNet real-world dataset demonstrate state-of-the-art 78.1 [email protected]:0.95, surpassing existing leaderboard systems with superior stability across diverse modulation bandwidths.

Figures & tables

Fig. 1: Overview of the proposed ZoomSpec architecture. The system strictly follows a Physics-Guided paradigm: the LS-STFT overcomes the geometric resolution bottleneck, while the AHLP module acts as a physics-guided operator to purify signal candidates before the Dual-Domain FRN.

Fig. 2: Visual comparison of spectral representations on simulated narrowband signals (Zigbee, LoRa, and NB-FM). While standard STFT suffers from sparse pixel support for narrowband emitters, the proposed LS-STFT significantly expands the visual footprint of signals like LoRa chirps and FM traces through periodic subband warping, enhancing feature prominence for detection.

Fig. 3: The AHLP processing chain. Guided by the coarse parameters ( $\hat{f}_{c}$ , $\hat{B}$ ) from the upstream CPN, the module performs baseband translation, bandwidth-matched adaptive filtering, and safe decimation to extract purified I/Q segments.

Fig. 4: FRN architecture. After AHLP, two streams-time-domain I/Q and FFT magnitude-pass through a 1-D downsampling stem and a shallow conv encoder, followed by a BiLSTM and two local-global additive-attention blocks to yield tokens $\mathbf{A}_{\mathrm{IQ}}$ and $\mathbf{A}_{\mathrm{FFT}}$ . A fusion bottleneck mixes channels to form fused tokens $\mathbf{F}$ . Heads specialize: Time Head consumes $\mathbf{A}_{\mathrm{IQ}}$ ; Bw Head consumes $\mathbf{A}_{\mathrm{FFT}}$ ; Class Head consumes $\mathbf{F}$ .

Dataset	SpaceNet
Modulation families	14
Sampling rate (MHz)	Train	5, 20, 30, 40, 50, 80
Sampling rate (MHz)	Test	20, 30, 40, 50, 80
Duration (ms)	Train	20, 40, 60, 80, 100, 150
Duration (ms)	Test	20, 40, 60, 80, 100, 200
Signals per file	Train	1-8

TABLE I: SpaceNet dataset acquisition settings and split.

Fig. 5: Class distribution of SpaceNet dataset.

Module	Parameter	Value	Physical/Empirical Justification
LS-STFT	$B_{\mathrm{sub}}$	1 MHz	Matches typical narrowband width and integer channel align.
LS-STFT	$\alpha_{1},\alpha_{2}$	1, 4	Prevents over-concentration while magnifying details.
AHLP	$\kappa$	0.2	Moderate bandwidth expansion for safety.
AHLP	$\eta$	0.1	Standard transition width (10%) for FIR design.
Training	Optimizer	AdamW	Better generalization stability.
	Learning Rate	$5\times 10^{-4}$	Standard for Transformer-based architectures.

TABLE II: Implementation details and physical justifications for hyperparameters.

Fig. 6: Visualization of the standard STFT spectrogram. Due to the linear frequency sampling, narrowband emissions (highlighted in the zoomed insets) occupy very few pixels on the frequency axis, resulting in weak contrast and blurred boundaries that impede small-object detection.

Fig. 7: Visualization of the proposed LS-STFT spectrogram under the same frequency budget $M$ . The periodic subband warping effectively increases the sampling density for narrowband components, significantly enhancing their visual prominence and sharpness for the detection network.

Method	mAP 50:95	Precision	Recall
YOLO11 STFT	50.3	0.76	0.62
YOLO11 LS-STFT	62.7	0.83	0.76
RF-DETR STFT	53.4	0.78	0.66
RF-DETR LS-STFT	69.9	0.86	0.81
D-FINE STFT	60.6	0.82	0.73
D-FINE LS-STFT	74.3	0.88	0.83

TABLE III: Detection performance comparison on SpaceNet dataset. Integrating the proposed LS-STFT consistently improves baseline detectors, while the full ZoomSpec framework achieves state-of-the-art results.

Fig. 8: mAP versus IoU threshold on SpaceNet dataset. Color indicates detector family; solid lines are LS-STFT and dashed lines are STFT.

Fig. 9: Per-class confusion matrices comparison on SpaceNet dataset. From left to right: RF-DETR ${}_{\text{LS-STFT}}$ , D-FINE ${}_{\text{LS-STFT}}$ , and the proposed method (Ours). The diagonal elements represent the classification recall, while off-diagonal elements indicate misclassification rates. Our method achieves the highest diagonal density across all categories, showing significant robustness in distinguishing narrowband signals (e.g., Zigbee, LoRa) and suppressing background confusion compared to the baselines.

Class	YOLO11		RF-DETR		D-FINE		Ours
Class	STFT	LS-STFT	STFT	LS-STFT	STFT	LS-STFT	Ours
WiFi 20MHz QPSK	66.1	66.2	71.2	70.5	70.3	71.2	71.5
WiFi 20MHz 16QAM	59.5	61.6	62.9	63.7	61.8	63.2	63.9
WiFi 20MHz 64QAM	65.8	66.6	70.1	72.3	72.5	72.8	72.8
WiFi 40MHz QPSK	60.1	73.4	64.2	78.3	71.7	78.4	78.4
WiFi 40MHz 16QAM	56.2	69.7	60.0	76.3	68.0	76.2	76.9

TABLE IV: Per-class [email protected]:0.95 comparison on SpaceNet dataset. The proposed LS-STFT input significantly boosts detection performance for narrowband and weak signals (e.g., Zigbee, LoRa, AM/FM) compared to standard STFT.

Method	Latency (ms)	Params (M)	Input size
YOLO11	11.3	59.3	$640\times 640$
RF-DETR	15.1	33.7	$640\times 640$
D-FINE	22.1	62.0	$640\times 640$
Ours	16.8	60.7	$640\times 640$

TABLE V: Model complexity and latency of compared methods on SpaceNet dataset. Latency is measured with batch size 1 on a single GPU.

Fig. 10: Confusion matrices for the 4-way bandwidth grading task in the CPN under different spectral representations. Standard STFT fails to resolve narrowband emissions, misclassifying 68.0% of them as background. The proposed LS-STFT significantly enhances narrowband visibility, boosting the recall rate from 31.0% to 89.0% while maintaining high accuracy for wideband signals.

Variant	LS-STFT	CPN	AHLP	FRN	Fusion Mode	[email protected]:0.95
Baseline with STFT	×	✓	×	✓	full	58.8
+ LS-STFT	✓	✓	×	✓	full	72.3
Fusion ablations within FRN
I/Q only	✓	✓	✓	✓	I/Q only	66.5
FFT only	✓	✓	✓	✓	FFT only	43.7
w/o cross-domain fusion	✓	✓	✓	✓	no fusion	76.7

TABLE VI: Ablation study on LS-STFT, CPN, AHLP, and FRN variants. mAP is evaluated as [email protected]:0.95.

Wireless Communications

Wireless communications studies how information is sent over radio links and how networks manage signals, spectrum, and resources reliably.

See all 399 papers

An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition
Aug 1, 2026Nurettin Safak, Durdu Can Yerdeyatar, Muhammet Sefa Demirel +3Automatic Modulation Classification
Department of Electrical and Electronics Engineering, Ankara Yildirim Beyazit University, Ankara, Turkey
RF-LEGO: Modularized Signal Processing-Deep Learning Co-Design for RF Sensing via Deep Unrolling
Apr 11, 2026Luca Jiang-Tao Yu, Chenshu WuDeep Unfolding Modular Neural Networks
The University of Hong Kong
Experimental Evaluation of a Low-Power Ultra-Wideband Receiver for Spectrum Sensing
Aug 24, 2026Panagiotis Vlachos, Ioannis A. Bartsiokas, Cedric Dehos +2Wireless Communications
IMS Laboratory, University of Bordeaux, CNRS, Bordeaux INP, IMS, UMR 5218, F-33400 Talence, France · Intracom Defense, Koropi, 19441, Greece · CEA-Leti, Universit´e Grenoble-Alpes, Minatec Campus Grenoble, France

Aug 1, 2026J/K move · Enter open · S save

An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition

Nurettin Safak, Durdu Can Yerdeyatar, Muhammet Sefa Demirel, Alperen Marasli+2

Department of Electrical and Electronics Engineering, Ankara Yildirim Beyazit University, Ankara, Turkey

Automatic Modulation Classification

Open paper AskPDF arXiv

Automatic RF modulation recognition is of critical importance in spectrum monitoring, electronic warfare, and cognitive radio applications, where low signal-to-noise ratio (SNR) conditions and the growing diversity of modulation schemes limit the performance of existing methods. This paper proposes an uncertainty-driven hybrid deep learning architecture for recognizing RF signals over a broad modulation space. The proposed approach carries out a multi-stage classification process by combining spectral information obtained through low-cost FFT-based preprocessing with time-frequency features extracted from short-time Fourier transform (STFT) spectrograms. The architecture comprises a 2D convolutional neural network (2D CNN)-based path for fast, low-latency primary classification, MC Dropout-supported Bayesian uncertainty estimation for assessing classification reliability, and a BiLSTM-based secondary decision mechanism activated under high-uncertainty conditions. The proposed system is evaluated in a controlled simulation environment spanning different SNR levels and modulation classes. Experimental results show that the primary 2D CNN path achieves

83.3\pm0.7\%

accuracy with an inference time of only 0.138 ms per sample, providing superior performance compared with traditional rule-based and classical machine-learning approaches. Furthermore, the obtained findings reveal the limitations of compact spectral feature representations and classifiers lacking temporal modeling, particularly in disambiguating FSK-based modulations. The uncertainty estimation module offers promising results for detecting low-confidence decisions, and the proposed approach demonstrates the potential of a low-latency and scalable solution for real-time RF modulation recognition.

Abstract

Wireless Communications

Explore similar work

An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition

RF-LEGO: Modularized Signal Processing-Deep Learning Co-Design for RF Sensing via Deep Unrolling

Experimental Evaluation of a Low-Power Ultra-Wideband Receiver for Spectrum Sensing