PaperlayerView exact revision on arXiv

Source-linked AI summary

SU-YOLO: Spiking Neural Network for Efficient Underwater Object Detection

Chenyang Li, Wenxuan Liu, Guoqiang Gong, Xiaobo Ding, Xian Zhong

arXiv:2503.24389v1cs.CVcs.NE

TL;DR

Underwater detection must handle difficult optical conditions while operating within equipment resource constraints, and lightweight models can trade accuracy for efficiency. SU-YOLO adapts YOLO as an SNN with denoising, SeBN, and redesigned spiking blocks; on URPC2019 it reports 78.8% mAP0.5 with 6.97M parameters and 2.98 mJ energy consumption. The authors note that energy-efficiency conclusions are estimates rather than hardware measurements.

  • Problem

    Underwater detection faces low-brightness, high-noise images and equipment size and power constraints, while lightweight models can sacrifice accuracy.

    Source: §1 · p. 2

  • Method

    SU-YOLO adapts YOLO as a lightweight SNN with redesigned spiking blocks, spike-based feature-map denoising, and SeBN.

    Source: §1 · pp. 3–4

  • Results

    78.8% mAP0.5 on URPC2019, with 6.97M parameters and 2.98 mJ energy consumption, surpasses existing lightweight SNNs and several ANN models.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The reported results demonstrate state-of-the-art performance among current SNN approaches for underwater object detection.

    Source: §5 · p. 27

    Energy-efficiency conclusions are estimates rather than actual hardware measurements.

    Source: §5 · p. 27

Abstract

from arXiv · show

Underwater object detection is critical for oceanic research and industrial safety inspections. However, the complex optical environment and the limited resources of underwater equipment pose significant challenges to achieving high accuracy and low power consumption. To address these issues, we propose Spiking Underwater YOLO (SU-YOLO), a Spiking Neural Network (SNN) model. Leveraging the lightweight and energy-efficient properties of SNNs, SU-YOLO incorporates a novel spike-based underwater image denoising method based solely on integer addition, which enhances the quality of feature maps with minimal computational overhead. In addition, we introduce Separated Batch Normalization (SeBN), a technique that normalizes feature maps independently across multiple time steps and is optimized for integration with residual structures to capture the temporal dynamics of SNNs more effectively. The redesigned spiking residual blocks integrate the Cross Stage Partial Network (CSPNet) with the YOLO architecture to mitigate spike degradation and enhance the model's feature extraction capabilities. Experimental results on URPC2019 underwater dataset demonstrate that SU-YOLO achieves mAP of 78.8% with 6.97M parameters and an energy consumption of 2.98 mJ, surpassing mainstream SNN models in both detection accuracy and computational efficiency. These results underscore the potential of SNNs for engineering applications. The code is available in https://github.com/lwxfight/snn-underwater.

1. Introduction

Underwater detection must contend with noisy, dim images and equipment constraints, while lightweight approaches can sacrifice accuracy. SU-YOLO adapts YOLO as an SNN with redesigned blocks, spike-based denoising, and SeBN; the reported URPC2019 result combines high detection accuracy with low energy consumption.

1. Introduction

  • Underwater detection is difficult because low brightness and high noise challenge image interpretation, while ROV and AUV size and power constraints require lightweight models.Lightweight models can reduce parameters and computational demands but typically lose accuracy relative to state-of-the-art general detectors.Source: §1 · p. 2
  • SU-YOLO is an SNN-based underwater detector built on YOLOv9's glean network and simplified for underwater-device energy constraints.Its lightweight SU-Blocks use CSPNet because conventional ResNet structures can cause spike degradation even in shallow networks.Source: §1 · p. 3
  • 78.8% mAP0.5 on URPC2019, with 6.97M parameters and 2.98 mJ energy consumption, exceeds existing lightweight SNNs and several ANN models.Figure 1 compares detection performance and energy consumption across models on URPC2019; blue dots are ANNs and green and red dots are SNNs.Source: §1 · p. 3, Figure 1 · p. 4
  • The spike-based denoising method improves feature-map quality using only integer additions and minimal computational overhead.The method processes feature maps to remove positive and negative noise.Source: §1 · pp. 3–4
  • SeBN concatenates feature maps across time steps along the channel dimension for batch normalization, with its design optimized for residual structures.The contribution summary reports that SeBN outperforms classical BN and tdBN.Source: §1 · p. 4
  • Experiments on URPC2019 and UDD are reported to validate the feasibility and performance of SNNs for underwater object detection.Source: §1 · p. 4

2. Related Work

Prior SNN detectors and underwater detection methods offer useful approaches, but the cited work leaves gaps in underwater-specific SNN optimization, resource demands, denoising compatibility, and normalization. SU-YOLO's stated design addresses these areas with a dedicated denoising method and SeBN.

2.1. SNN Object Detection

  • Conversion-based SNN detectors reduced time steps from more than 1000 to 300 or, with Fast-SNN, 3, while direct-training methods use surrogate gradients to address non-differentiable firing.Fast-SNN's negative-spike computation compromises binary spikes; underwater-focused SNN optimization remains limited because EMS-YOLO and SpikeYOLO target general detection and perform suboptimally in blurry, noisy underwater scenes.Source: §2.1 · p. 5

2.2. Underwater Object Detection

  • Two-stage underwater detectors can incur high computational costs, while attention-based YOLO variants increase computational load and lightweight ULO reports 65.13% mAP on URPC.Single-stage YOLO methods are described as simpler and more suitable for resource-constrained underwater devices; SU-YOLO is proposed in response to the accuracy and resource trade-offs.Source: §2.2 · pp. 5–6

2.3. Underwater Image Denoising

  • Conventional and deep-learning denoisers can blur details, require intensive floating-point computation, or run as separate models, motivating a denoising method designed for SNNs.The cited approaches include filtering, enhancement, wavelet, Transformer-CNN, and diffusion methods.Source: §2.3 · pp. 6–7

2.4. Normalization

  • Existing SNN normalization methods either merge time steps or normalize them independently, but the latter approaches cited do not account for shortcut connections or batch-scale fusion.The authors report better performance from independently normalizing each time step than from merging features across time steps, and propose SeBN to address the remaining issues.Source: §2.4 · p. 7

3. Proposed Method

SU-YOLO combines spiking YOLO modules for underwater detection, processing input through encoding, denoising, feature extraction, multi-scale processing, feature fusion, and prediction. Its components address noisy binary feature maps, spike degradation, and time-dependent normalization.

3.1. Spiking Underwater YOLO

  • The network feeds normalized RGB images to an Encoder, then processes features through SpikeDenoiser, three SU-Block1 backbone blocks, SpikeSPP, neck fusion, and DetectHead.SU-YOLO uses IF neurons for its spiking neurons; unlike LIF neurons, they have no membrane-voltage decay over time.Source: §3.1 · pp. 7–8

3.2. SpikeDenoiser

  • The SpikeDenoiser sequentially applies three convolution filters to spike feature maps, inverting a pixel when the convolution result equals 4.Kernels K1 and K2 fill missing pixels; K3 removes isolated pixels.Source: §3.2 · p. 9
  • 3.5% mAP improvement on UDD; SpikeDenoiser uses integer additions at inference, costing less than 0.1 GSOPs and 0.01 mJ per 320 × 320 image.It fills missing pixels and removes isolated pixels while keeping input and output feature maps binary.Source: §3.2 · p. 10

3.3. SU-Block

  • SU-Block1 integrates CSPNet and shortcut connections for backbone feature extraction, downsampling, and channel expansion; SU-Block2 supports neck feature fusion without downsampling or extra residual connections.CSPNet mitigates the depth-related spike degradation described for ResNet and maintains neuron activity through the backbone.Source: §3.3 · pp. 11–12

3.4. SpikeSPP

  • SpikeSPP pools feature maps before activation to better preserve information in binary-spike SNNs, and removing its third max-pooling layer reduces computation without sacrificing performance.Its resulting features pass to IF neurons, which generate spikes with appropriate firing rates.Source: §3.4 · p. 12

3.5. Encoder and DetectHead

  • The Encoder repeats a conventionally extracted and normalized feature map across T time steps and converts it into binary spikes; the DetectHead predicts classes and boxes from fused multi-level features.The DetectHead’s final spiking neuron accumulates floating-point values over time with an infinite membrane-potential threshold.Source: §3.5 · p. 13

3.6. Separated Batch Normalization

  • SeBN normalizes feature maps independently across time steps and is optimized for residual structures, preserving temporal differences while addressing distribution adjustment before residual summation.At inference, batch-scale fusion removes the normalization layer, while dedicated parameters for each time step increase model file size without changing computational load or inference time.Source: §3.6 · pp. 13, 15, 16

4.1. Datasets and Implementation Details

The experiments use underwater detection datasets and a fixed training setup, and assess SNN computational cost with synaptic operations alongside time-step effects. URPC2019 and UDD provide the underwater evaluation data, while Pascal VOC 2012 is used for SeBN experiments.

4.1.1. Datasets

  • URPC2019 contains 4,707 labeled images across four classes, and UDD contains 2,227 labeled images across three classes.URPC2019 is split into 3,767 training, 695 validation, and 245 test images; UDD has 1,827 training and 400 validation images.Source: §4.1.1 · p. 16
  • Pascal VOC 2012 contributes 11,540 labeled images across 20 classes for experiments involving SeBN.The dataset is split into 5,717 training and 5,823 validation images.Source: §4.1.1 · p. 16

4.1.2. Implementation Details

  • Training uses IF neurons, SGD at a 0.001 learning rate, batch size 16, 320 × 320 inputs, and 300 epochs.Experiments run on one NVIDIA RTX 2080Ti GPU with an AMD Ryzen 5600 CPU and 32GB RAM.Source: §4.1.2 · p. 16

4.1.3. Computational Cost

  • SOPs measure SNN computational cost by accounting for convolution operations, preceding-layer spike emissions, and additions during neuron charging.The paper estimates energy using 4.6 pJ per MAC and 0.9 pJ per addition on a typical 45nm chip.Source: §4.1.3 · p. 17

4.1.4. Time Step

  • More time steps increase spike emissions and expressiveness but add computational overhead and reduce inference speed.For equivalent detection accuracy, fewer time steps are more advantageous for energy efficiency and real-time performance.Source: §4.1.4 · pp. 17–18

4.2. Comparison Experiments

On URPC2019 and UDD, SU-YOLO is compared with converted and directly trained SNNs as well as lightweight ANNs. It achieves higher accuracy than compared SNNs with fewer parameters and time steps, while its energy use remains low relative to lightweight ANN comparisons.

4.2. Comparison Experiments

  • Across URPC2019 and UDD, the directly trained SNN outperforms converted SNNs using fewer time steps, and exceeds directly trained SNNs in mAP with fewer parameters.Its performance is reported as comparable to lightweight ANNs.Source: §4.2 · p. 18
  • SU-YOLO’s mAP is slightly below YOLOv7’s, but its energy consumption is 5% of YOLOv7’s.Against lightweight ANNs with similar energy profiles, SU-YOLO demonstrates superior performance.Source: §4.2 · pp. 18–19
  • 0.788 mAP0.5, 6.97M parameters, and 2.98 mJ energy consumption: SU-YOLO surpasses EMS-YOLO and SpikeYOLO on URPC2019.EMS-YOLO reports mAP0.5 of 0.672 at 2.73 mJ for ResNet10 and 0.735 for ResNet34; SpikeYOLO reports 0.784 at 8.64 mJ.Source: §4.2 · p. 18, Table 1 · p. 19

4.3. Ablation Experiments

The ablations evaluate SU-YOLO’s modules and design choices across detection accuracy, feature-map behavior, and computational load. Together, they report gains from the combined modules, denoising, normalization, residual-block replacement, pooling order, and a four-step setting.

4.3.1. Effectiveness of Modules in SU-YOLO

  • Applying all proposed modules raises mAP0.5 by 4.5% and mAP0.5:0.95 by 3.3%; SpikeDenoiser alone adds 0.9% mAP0.5 over baseline.Each module incrementally reduced parameter count and improved accuracy.Source: Table 2 · p. 20, §4.3.1 · p. 20

4.3.2. Effectiveness of SpikeDenoiser

  • SpikeDenoiser improves mAP on both datasets with minimal added energy, while the Mean Filter degrades mAP and the Gaussian Filter slightly improves accuracy.The visualization reports that SpikeDenoiser removes noise from feature maps and provides clearer inputs to the backbone.Source: Table 3 · p. 20, §4.3.2 · pp. 20–21, Figure 9 · p. 21

4.3.3. Effectiveness of SeBN

  • SeBN improves detection accuracy over BN alternatives and achieves the highest mAP in tests on EMS-YOLO and Pascal VOC 2012.Its feature distributions remain stable and close to N(0, 1) at each time step, unlike BN and TEBN fluctuations and tdBN’s variance deviation.Source: Table 4 · p. 21, §4.3.3 · pp. 21–22, Figure 10 · p. 22

4.3.4. Residual Block Replacement

  • SU-Blocks yield firing rates of 12.7%, 15.4%, and 18.8%, versus EMS-Blocks’ 15.2%, 11.9%, and 11.5%, and SU-YOLO reaches 78.8% mAP0.5.The authors associate the CSPNet-based SU-Blocks with mitigating spike degradation and enhancing feature extraction.Source: Table 5 · p. 23, §4.3.4 · p. 23

4.3.5. Confusion Matrix and Precision-Recall Curve

  • SU-Block achieves 78.8% mAP0.5, compared with 75.4% for ANN-Res18 and 72.2% for EMS-Block, with particular gains in scallop detection.The confusion-matrix discussion reports scallop accuracy advantages of 12% and 16% over ANN-Res18 and EMS-Res18, respectively.Source: Figure 12 · p. 24, §4.3.5 · p. 24, Figure 13 · p. 25

4.3.6. Sequence of Pooling and Activation

4.3.7. Length of Time Steps

  • SU-YOLO achieves optimal accuracy at T = 4; increasing T beyond 4 does not improve performance and may slightly reduce it.The authors adopt four time steps as the balance between detection accuracy and power consumption.Source: Table 7 · p. 26, §4.3.7 · p. 26

5. Conclusion

The conclusion presents SU-YOLO as a lightweight underwater SNN with effective denoising and SeBN, reporting state-of-the-art performance among current underwater SNNs. Its energy-efficiency conclusions remain estimates, and broader evaluation and hardware implementation are future work.

5. Conclusion

  • The proposed model combines spike-based image denoising and SeBN, with carefully designed modules reported to enhance underwater detection capabilities.The conclusion says the denoising method proved highly effective in experiments and SeBN outperformed existing normalization techniques.Source: §5 · p. 26
  • SU-YOLO achieves state-of-the-art performance among current SNN approaches for underwater object detection.The authors also describe its denoising method as effective and SeBN as outperforming existing normalization techniques.Source: §5 · pp. 26–27
  • Energy-efficiency conclusions are based on estimates rather than actual hardware measurements; hardware implementation and evaluation on additional datasets are future work.The authors also plan to enhance performance and explore broader object-detection tasks and applications.Source: §5 · p. 27

Author Biography

The biographies describe five researchers' education, current roles, and research interests. Their interests include computer vision, neuromorphic computing, spike vision, action recognition, multimedia content analysis, intelligent signal processing, IoT technology, and system virtualization.

Author Biography

  • Chenyang Li is a master's student in computer technology at China Three Gorges University, with research interests in computer vision and neuromorphic computing.He received his bachelor's degree in water conservancy and hydropower engineering from the same university in 2021.Source: Author Biography · p. 33
  • Wenxuan Liu received her Ph.D. from Wuhan University of Technology in 2024 and is a postdoc at Peking University.Her research interests include neuromorphic computing, spike vision, action recognition, and multimedia content analysis.Source: Author Biography · p. 33
  • Guoqiang Gong is a professor at China Three Gorges University whose research interests include intelligent signal processing and computer vision.He received his bachelor's degree from Lanzhou University in 1998 and his Ph.D. from Tongji University in 2010.Source: Author Biography · p. 33
  • Xiaobo Ding is an associate professor at China Three Gorges University with research interests in IoT technology and system virtualization.The biography lists bachelor's and Ph.D. degrees from Beijing Electronic Science and Technology Institute and Harbin Engineering University, respectively.Source: Author Biography · p. 33
  • Xian Zhong is a professor at Wuhan University of Technology whose research interests include multimodal information processing, intelligent transportation systems, and neuromorphic computing.He received his bachelor's and Ph.D. degrees in computer science from Wuhan University and Huazhong University of Science and Technology, respectively.Source: Author Biography · p. 33

Figure Captions

The supplied caption text defines k in relation to the mean and variance of the kth feature map.

Figure Captions

  • k denotes the mean and variance of the kth feature map.The supplied caption fragment gives no further figure or comparison information.Source: Figure Captions · p. 34
Loading 2503.24389v1…