Spiking Neural Networks for Energy-Efficient Object Detection in Forward-Looking Sonar Imagery

Gwenevere Frank; Gert Cauwenberghs

cs.CVAug 22, 2026

Spiking Neural Networks for Energy-Efficient Object Detection in Forward-Looking Sonar Imagery

Authors: Gwenevere Frank, Gert Cauwenberghs

Organizations: Department of Electrical Engineering, University of California San Diego, La Jolla, CA 92093, USA · Department of Bioengineering, University of California San Diego, La Jolla, CA 92093, USA

Abstract

Autonomous underwater vehicles (AUVs) are increasingly important tools in industries ranging from research, to energy, to defense. AUVs are power-constrained platforms operating in remote environments with fixed battery capacities, where propulsion competes with compute and sensors for power over lengthy mission durations. AUVs frequently operate in dark or turbid waters where optical sensing is of limited value, and rely on sonar as their primary sensing modality. Convolutional neural networks (CNNs) are the state-of-the-art solution for object detection in forward-looking sonar imagery, but are energy expensive (e.g. YOLOv8m: 322 mJ/inference). Spiking neural networks (SNNs) rely on binary spike activations and thus sparse accumulate-only operations, allowing them to be remarkably energy efficient, particularly when paired with dedicated neuromorphic hardware. The sparse, high-contrast structure of forward-looking sonar (FLS) returns is structurally matched to spike coding in a way that optical imagery is not. No prior work has assessed the suitability of SNNs for object detection in FLS imagery. SpikeYOLO, a fully spiking network trained with surrogate gradients, was benchmarked against state-of-the-art CNN baselines on three FLS object detection datasets. Key results: SpikeYOLO T=2 achieves 3.3 $\times$ lower theoretical compute energy on UATD (97 vs 322 mJ) at competitive accuracy (0.529 [email protected]:0.95 vs. YOLOv8m's 0.575); SpikeYOLO matches YOLOv8m on [email protected] and outperforms YOLO-SONAR and Fast R-CNN baselines on the sparse Marine-Debris-FLS dataset at 4.4 $\times$ lower energy; SpikeYOLO demonstrates superior robustness to multiplicative speckle noise (3.0% degradation at $σ{=}0.4$ vs. 8.9% for YOLOv8m), outperforming YOLOv8m outright at $σ{=}0.6$ , directly relevant to real-world FLS deployment.

Figures & tables

Figure 1: Representative FLS images from the three evaluation datasets. Top row: UATD (10 shape classes, fine-grained discrimination). Middle row: Marine-Debris-FLS (10 classes, sparse, high-contrast watertank returns). Bottom row: WHFLS (3 classes, victim/boat/plane). Ground-truth bounding boxes shown in color per dataset.

Operation	8-bit	16-bit	32-bit
Integer add	0.03 pJ	—	0.1 pJ
Integer multiply	0.2 pJ	—	3.1 pJ
FP add	—	0.4 pJ	0.9 pJ
FP multiply	—	1.1 pJ	3.7 pJ

Table 1: Operation energy costs in 45 nm CMOS at 0.9 V [ 14 ] .

Figure 2: SpikeYOLO [ 1 ] and YOLOv8m [ 12 ] training and evaluation pipeline. Both models are initialized from a COCO checkpoint and fine-tuned on each FLS dataset.

Model	Init. checkpoint	Optimizer / LR / schedule	Epochs
YOLOv8m [ 12 ]	COCO	SGD, lr = 0.01, cosine	50
RT-DETR-L [ 9 ]	COCO	Ultralytics defaults	50
SpikeYOLO [ 1 ]	COCO ( $T{=}4$ )	AdamW, lr = $10^{-4}$ , cosine	50

Table 2: Training configuration for all retrained models.

Model	Type	[email protected]	[email protected]:0.95	Energy (mJ)
Faster R-CNN [ 10 ]	CNN	0.839	—	—
RT-DETR-L [ 9 ]	Transformer	0.979	$0.642\pm 0.005$	476
YOLOv8m [ 12 ]	CNN	0.969	$0.575\pm 0.004$	322
SpikeYOLO (T=2) [ 1 ]	Fully spiking	0.964	$0.529\pm 0.005$	97

Table 3: UATD detection results. Bold : best per column. Dashes: values not published in literature for architecture. All [email protected]:0.95 values are means over 3 independent seeds.

Figure 3: UATD detection accuracy (mAP@0.5 and mAP@0.5:0.95) by network. SNN achieves competitive accuracy. — Figure 3: UATD detection accuracy ([email protected] and [email protected]:0.95) by network. SNN achieves competitive accuracy.

Model	Type	[email protected]	[email protected]:0.95	Energy (mJ)
YOLO-SONAR [ 19 ]	CNN	0.953	0.667	—
Fast R-CNN [ 11 ]	CNN	0.967	0.759	—
YOLOv8m [ 12 ]	CNN	0.987	$0.801\pm 0.006$	322
RT-DETR-L [ 9 ]	Transformer	0.989	$0.819\pm 0.008$	476
SpikeYOLO (T=2) [ 1 ]	Fully spiking	0.987	$0.781\pm 0.008$	73

Table 4: Marine-debris-fls detection results. Bold : best per column. Dashes: values not published in literature for architecture. All [email protected]:0.95 values are means over 3 independent seeds.

Figure 4: Marine-debris-fls detection accuracy (mAP@0.5 and mAP@0.5:0.95) by network. SpikeYOLO matches YOLOv8m on mAP@0.5. — Figure 4: Marine-debris-fls detection accuracy ([email protected] and [email protected]:0.95) by network. SpikeYOLO matches YOLOv8m on [email protected].

Model	Type	[email protected]	[email protected]:0.95	Energy (mJ)
YOLOv8m [ 12 ]	CNN	0.9949	$0.872\pm 0.004$	322
RT-DETR-L [ 9 ]	Transformer	0.9950	$0.893\pm 0.001$	476
SpikeYOLO (T=2) [ 1 ]	Fully spiking	0.9950	$0.861\pm 0.001$	52

Table 5: WHFLS detection results. Bold : best per column. Dashes: values not published in literature for architecture. All [email protected]:0.95 values are means over 3 independent seeds.

Figure 5: WHFLS detection accuracy (mAP@0.5 and mAP@0.5:0.95) by network. SpikeYOLO beats YOLOv8m on mAP@0.5 (0.9950). — Figure 5: WHFLS detection accuracy ([email protected] and [email protected]:0.95) by network. SpikeYOLO beats YOLOv8m on [email protected] (0.9950).

Figure 6: Inference energy breakdown per model. SpikeYOLO uses accumulate-only operations, achieving 3.3 $\times$ lower theoretical energy than YOLOv8m at $T{=}2$ .

Model	Params (M)	Energy (mJ)	[email protected]:0.95
YOLOv8m [ 12 ]	25.9	322	$0.575\pm 0.004$
SpikeYOLO (T=2) [ 1 ]	23.1	97	$0.529\pm 0.005$

Table 6: Model efficiency comparison on UATD. Energy: theoretical CMOS estimate (

E_{\text{MAC}}=

4.6\text{,}\mathrm{p}\mathrm{J}

, $E_{\text{AC}}=$0.9\text{\,}\mathrm{p}\mathrm{J}

, [ 14 ] ).

Model	Dataset	$\bar{s}$
SpikeYOLO T=2	UATD	0.880
	Marine Debris	0.668
	WHFLS	0.476

Table 7: Mean spike activation

\bar{s}

(mean I-LIF integer output per neuron per timestep, range 0–4) measured on each FLS test set at

T{=}2

Figure 7: mAP@0.5:0.95 vs. inference timesteps $T$ for SpikeYOLO on all three FLS datasets. Accuracy is relatively flat across $T\in\{2,4,8\}$ ; energy scales linearly with $T$ , $T{=}2$ is Pareto-optimal. — Figure 7: [email protected]:0.95 vs. inference timesteps $T$ for SpikeYOLO on all three FLS datasets. Accuracy is relatively flat across $T\in\{2,4,8\}$ ; energy scales linearly with $T$ , $T{=}2$ is Pareto-optimal.

Figure 8: mAP@0.5:0.95 under multiplicative speckle noise, evaluated on UATD test set without retraining. SpikeYOLO [ 1 ] degrades 3.0% at $\sigma{=}0.4$ vs. 8.9% for YOLOv8m [ 12 ] . At $\sigma{=}0.6$ the fully-spiking model outperforms the ANN baseline outright. — Figure 8: [email protected]:0.95 under multiplicative speckle noise, evaluated on UATD test set without retraining. SpikeYOLO [ 1 ] degrades 3.0% at $\sigma{=}0.4$ vs. 8.9% for YOLOv8m [ 12 ] . At $\sigma{=}0.6$ the fully-spiking model outperforms the ANN baseline outright.

Training fraction	YOLOv8m [ 12 ]	SpikeYOLO T=2 [ 1 ]
25%	$0.449\pm 0.011$	$\mathbf{0.475}\pm 0.006$
50%	$0.491\pm 0.009$	$\mathbf{0.513}\pm 0.005$
100%	$0.575\pm 0.004$	$0.529\pm 0.005$

Table 8: Low-data ablation: [email protected]:0.95 on UATD test set at reduced training fractions. Mean

\pm

std over 2 random seeds. 100% results from 3-seed retrains (Section 4.1 ).

Figure 9: mAP@0.5:0.95 vs. training data fraction for YOLOv8m [ 12 ] and SpikeYOLO [ 1 ] T=2 on UATD [ 10 ] . SpikeYOLO leads at 25% and 50% data; YOLOv8m recovers at 100%. — Figure 9: [email protected]:0.95 vs. training data fraction for YOLOv8m [ 12 ] and SpikeYOLO [ 1 ] T=2 on UATD [ 10 ] . SpikeYOLO leads at 25% and 50% data; YOLOv8m recovers at 100%.

Figure 10: Per-class mAP@0.5:0.95 on UATD [ 10 ] . Plane and ball achieve the highest accuracy across all models; square cage and cylinder are consistently challenging due to complex geometry. — Figure 10: Per-class [email protected]:0.95 on UATD [ 10 ] . Plane and ball achieve the highest accuracy across all models; square cage and cylinder are consistently challenging due to complex geometry.

Figure 11: Qualitative detection results on UATD [ 10 ] test images. Left: SpikeYOLO [ 1 ] T=2. Right: YOLOv8m [ 12 ] . Dashed white: ground truth. Solid: predictions with class label and confidence score. Rows: plane, cylinder, circle cage, ROV.

Dataset	compute	+DRAM	compute	+DRAM	compute-only	+DRAM
	SpikeYOLO (mJ)		YOLOv8m (mJ)		Ratio
UATD	97	112–127	322	339–356	3.3 $\times$	2.8–3.0 $\times$
Marine-Debris-FLS	73	88–103	322	339–356	4.4 $\times$	3.5–3.9 $\times$
WHFLS	52	67–82	322	339–356	6.2 $\times$	4.3–5.1 $\times$

Table 9: SpikeYOLO vs. YOLOv8m energy, compute-only vs. including the estimated single-read DRAM weight-fetch overhead (15–34 mJ, fixed by parameter count rather than dataset).

Efficient Object Detection

Efficient object detection reduces the computation, latency, memory, and energy needed to find and locate objects in images and video.

See all 95 papers

Apr 30, 2026

Real-Time Frame- and Event-based Object Detection with Spiking Neural Networks on Edge Neuromorphic Hardware: Design, Deployment and Benchmark

Real-time object detection on energy-constrained platforms is critical for applications such as UAV-based inspection, autonomous navigation, and mobile robotics. Spiking neural networks (SNNs) on neuromorphic hardware are believed to be significantly more energy-efficient than conventional artificial neural networks (ANNs). In this work, we present a comprehensive methodology for designing general SNN detection architectures targeting neuromorphic platforms, along with the engineering adaptations required to deploy them on the state-of-the-art Neuromorphic processor, Intel Loihi 2. We benchmark SNN-based object detection on Loihi 2 using both frame-based and event-based datasets, comparing performance with ANN-based detection on the NVIDIA Jetson Orin Nano, NVIDIA Jetson Nano B01, and the Apple M2 CPU. Our results show that SNNs on Loihi 2 can perform real-time detection while achieving the lowest per-inference dynamic energy among all platforms. Also, Loihi 2 outperforms the other platforms in terms of power consumption, though ANNs on Jetson Orin Nano achieve higher inference rates. Furthermore, our ANN-to-SNN distillation-aware training enables SNNs to recover 87-100% of the detection accuracy of their ANN counterparts while maintaining lower inference latency; without distillation, SNNs exhibit an 11-27% accuracy drop. These results highlight the potential of neuromorphic systems for energy-efficient, real-time object detection at the edge.

Udayanga G. W. K. N. Gamage, Yan Zeng, Cesar Cadena +2

Department of Electrical and Photonics Engineering, Technical University of Denmark, Anker Engelunds Vej 1, 2800 Kongens Lyngby, Denmark · Faculty of Engineering, University of Ruhuna, Hapugala,Galle 80000, Sri Lanka · Autonomous Systems Lab, ETH zürich, Leonhardstrasse 21, 8092 Zürich, Switzerland

Edge Inference Spiking Neural Networks Neuromorphic Computing Event-Based Object Detection Real-Time Object Detection

Ask PDF

May 24, 2026

MambaDSF: Multi-Scale SSM with Dilated Feature Fusion for Sonar Small Target Detection

Sonar imaging is the primary modality for underwater target detection, yet small targets remain difficult to detect due to insufficient pixel coverage, low acoustic contrast, and scale ambiguity across imaging ranges. CNN-based detectors extract local features efficiently but cannot suppress noise-induced false alarms without global acoustic context. Transformer-based methods capture long-range dependencies at quadratic computational cost. Existing Mamba-based vision models offer efficient linear-cost scanning but lack multi-scale semantic alignment across pyramid levels, multi-receptive-field fusion, and small-target-aware training supervision needed for reliable sonar detection. This letter proposes Mamba Dilated-Scale Fusion (MambaDSF), a hybrid framework addressing these limitations through three contributions: a Mamba Enhanced Feature Pyramid (MambaEFP) backbone that jointly captures local echo cues and global acoustic context at linear complexity; a Dilate Fusion Mamba (DFMamba) encoder that enforces multi-scale feature alignment across pyramid levels; and Scale-Adaptive Weighted IoU (SA-WIoU) and Cross-Scale Coherence (CSC) losses that stabilize small-target training. MambaDSF achieves 91.5% mAP50 on the UATD forward-looking sonar benchmark with 28.7 million parameters, surpassing all compared detectors. On a small-target subset the gain reached +2.2 percentage points, and cross-domain evaluation on FLS and MD-FLS confirms the generalization of the proposed architecture. The codes are publicly available at https://github.com/IDontKnowAAA/MambaDSF.

Hui Lin, Jiayi Li, Jing Wang +1

School of Information Science and Engineering, Ocean University of China, Qingdao 266100, China · School of Information and Communication Technology, Griffith University, Nathan, QLD 4111, Australia

Underwater Imaging Small Object Detection Multi-Scale Feature Fusion Visual SSMs Underwater Object Detection

Ask PDF

Mar 29, 2026

STATrack: A Target-Aware Fully Spiking Neural Network for Efficient UAV Tracking

Spiking Neural Networks (SNNs), characterized by their event-driven computation and low power consumption, have shown great potential for energy-efficient visual tracking on unmanned aerial vehicles (UAVs). However, existing SNN-based trackers often rely on costly event cameras, which limits their deployment on standard RGB-camera UAV platforms. To address this limitation, we propose STATrack, a fully spiking neural network framework for UAV visual tracking using only RGB inputs. To the best of our knowledge, this is the first study to explore fully spiking neural networks for RGB-based UAV visual tracking. Considering the challenge of preserving fine-grained target information in deep spiking representations under a limited number of time steps, we introduce Adaptive Mutual Information Maximization (AMIM). AMIM maximizes the statistical dependency between target-centered templates and their deep spiking representations, encouraging the spiking backbone to learn representations that better distinguish the target from surrounding backgrounds. We further develop a sample-difficulty-aware dynamic weighting strategy that adaptively adjusts the mutual-information constraint according to the relative localization difficulty of each training batch. Extensive experiments on four widely used UAV tracking benchmarks demonstrate that STATrack achieves state-of-the-art tracking performance with low theoretical energy consumption, highlighting its potential for energy-constrained UAV applications. Code is released at: https://anonymous.4open.science/r/STATrack.

Pengzhi Zhong, Jiwei Mo, Dan Zeng +2

Guilin University of Technology, Guilin 541006, China · School of Artificial Intelligence, Sun Yat-sen University, Zhuhai 510275, China · School of Electronic Information, Central South University, Changsha, Hunan, 410083, China

Visual Object Tracking Spiking Neural Networks

Ask PDF

Abstract

Efficient Object Detection

Explore similar work

Real-Time Frame- and Event-based Object Detection with Spiking Neural Networks on Edge Neuromorphic Hardware: Design, Deployment and Benchmark

MambaDSF: Multi-Scale SSM with Dilated Feature Fusion for Sonar Small Target Detection

STATrack: A Target-Aware Fully Spiking Neural Network for Efficient UAV Tracking