An Integrated Lightweight YOLO Framework Combining Structured Pruning, Knowledge Distillation, and Multi-Frame Linear Interpolation for Robust Real-Time Nonlinear Dynamic Response Analysis of Small-Scale Underwater Targets under Bubble Loading

Abstract

A You Only Look Once version 8 (YOLOv8)-based detection model optimized via pruning and knowledge distillation is proposed for the challenging task of detecting dynamic responses of small underwater bubbles. Underwater bubble experiments were conducted to collect dynamic image sequences, and the improved YOLOv8 model was used to detect bubbles and extract their nonlinear dynamic responses. Through model pruning and distillation, the optimized model considerably reduces model parameters and accelerates inference while maintaining high detection accuracy. The experimental detection results were validated by a coupled Eulerian-Lagrangian computational fluid dynamics simulation, which showed high consistency with the observed bubble dynamics. The proposed approach demonstrates excellent performance in recognizing small targets with nonlinear dynamic responses, shows strong robustness in complex underwater scenarios, and offers high inference efficiency and ease of deployment.

Keywords

PruneDistill-YOLOv8 / Structured pruning / Knowledge distillation / Underwater target detection / Bubble dynamics

References

[1] AwangMK, AminuddinHS, KamilNKM, MustafaKA. Class 3 wiener filtering for underwater image enhancement and restoration. 2021 International Conference on Engineering and Emerging Technologies (ICEET), 2021, Piscataway. IEEE. 489494

[2] ChaffreT, SantosPE, LeC G, ChauveauE, SammutK, ClementB. Learning adaptive control of a uuv using a bioinspired experience replay mechanism. IEEE Access, 2023, 11: 123505-123518

[3] ChengC, HouX, WangC, WenX, LiuW, ZhangF. A pruning and distillation based compression method for sonar image detection models. Journal of Marine Science and Engineering, 2024, 1261033

[4] ChoC, JinYS. Rotary trigger double spark gap switch for repetitive pulse generator and production of silver nanocolloid. Review of Scientific Instruments, 2018, 898084707

[5] DengY, GuoZN, HuangZG. Simulation of laser-induced asymmetric bubble with lattice boltzmann method. Advanced Materials Research, 2012, 591-593: 806-809

[6] ElmezainM, SaadS L, SultanA, HeshmatM, SeneviratneL, HussainI. Advancing underwater vision: a survey of deep learning models for underwater object recognition and tracking. IEEE Access, 2025, 13: 17830-17867

[7] FangS, ChiC, LiuD, DingJ, WangK, ShiY, ZhangC, TangG, ChenX. The development of a high speed underwater optoelectronic imaging module. Lidar Imaging Detection and Target Recognition, 2017, Bellingham. SPIE. 10605106550V

[8] GeS, LuoZ, ZhangC, HuaY, TaoD. Distilling channels for efficient deep tracking. IEEE Transactions on Image Processing, 2020, 29: 2610-2621

[9] HanW, XingY, LiR, LiW, HaoY, ChenY. Dynamic characteristics of near-wall single bubble collapse-induced shock waves. Journal of Applied Fluid Mechanics, 2025, 183: 617-630

[10] HaoS, LiJ, SunX, MaX, AnB, HeT. MDFOaNet: A novel multi-modal pedestrian detection network based on multi-scale image dynamic feature optimization and attention mapping. IEEE Transactions on Intelligent Transportation Systems, 2025, 261: 268-282

[11] HartmannA, KlaasM, SchroederW. Coupled airfoil heave/pitch oscillations at buffet flow. AIAA Journal, 2013, 517: 1542-1552

[12] HeGH, KashiwagiM. Numerical analysis of the hydroelastic behavior of a vertical plate due to solitary waves. Journal of Marine Science and Technology, 2012, 172: 154-167

[13] HoYJ, ChangHC, LinCW, FanCH, LinYC, WeiKC, YehCK. Oscillatory behavior of microbubbles impacts efficacy of cellular drug delivery. Journal of Controlled Release, 2021, 333: 316-327

[14] HongY, LiuY, YangS, ZhangK, HuJ. Joint extraction of entities and relations using graph convolution over pruned dependency trees. Neurocomputing, 2020, 411: 302-312

[15] HuX, KeY. EMFANet: a lightweight network with efficient multi-scale feature aggregation for real-time semantic segmentation. Journal of Real-time Image Processing, 2024, 212: article nu-ber 40

[16] JinX, ZhangD, WuQ, XiaoX, ZhaoP, ZhengZ. Improved SiamCAR with ranking-based pruning and optimization for efficient UAV tracking. Image and Vision Computing, 2024, 141104886

[17] JosephKJ, RajasegaranJ, KhanS, KhanFS, BalasubramanianVN. Incremental object detection via meta-learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 4412: 9209-9216

[18] KalatehF, KooshehA. Finite element analysis of flexible structure and cavitating nonlinear acoustic fluid interaction under shock wave loading. International Journal of Nonlinear Sciences and Numerical Simulation, 2018, 195: 459-473

[19] KaytoueM, PlantevitM, ZimmermannA, BendimeradA, RobardetC. Exceptional contextual subgraph mining. Machine Learning, 2017, 1068: 1171-1211

[20] KazantzidouC, PerezT, ValentinisF, DonaireA. On the tuning of a nonlinear energy-based regulator for the positioning of a fully actuated surface marine craft. IET Control Theory and Applications, 2021, 156: 850-860

[21] KedrinskiiVK. Multiphase systems: Wave processes, state dynamics, and mathematical models. Journal of Physics: Conference Series, 2016, 754042007

[22] LiC, TengX, DingY, LanL. Instance-level scaling and dynamic margin-alignment knowledge distillation for remote sensing image scene classification. Remote sensing, 2024, 16203853

[23] LiJB, XuWL, ZhaiYW, LuoJ, WuH, DengJ. Experimental study of the collapse strength of a bubble merged from a sparkinduced bubble and a pre-existing air bubble. Ocean Engineering, 2022, 249110907

[24] LiS, SaadeY, vand M D, LohseD. Comparison of boundary integral and volume-of-fluid methods for compressible bubble dynamics. International Journal of Multiphase Flow, 2021, 145103834

[25] LiangW, LinY, ZhuC, ZhangY, ChenR. Experimental study on dynamic response of aluminum film under two spark bubbles coupling loading. Ocean Engineering, 2022, 266Part3112888

[26] LinMG, HuangCT, ChuangYC, ChenYT, HsuYT, ChenYK, ChouJJ, LiuTT, ShihCS, WuAY. D-NAT: Data-driven nonideality aware training framework for fabricated computingin-memory macros. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2022, 122: 381-392

[27] LinX, SunY, LiuY, JiangK, CaoA. Stabilization of highenergy cathode materials of metal-ion batteries: control strategies and synthesis protocols. Energy & Fuels, 2021, 359: 7511-7527

[28] LiuF, XuJ, ZhangL, DuR, SuZ, ZhangA, HuZ. DESNN algorithm for communication network intrusion detection. Wireless Personal Communications, 2022, 1262: 1705-1720

[29] Liu H, Ye X, Zhou H, Huang H (2025) Research on UUV carrying forward-looking sonar for target location based on spatial analysis. IEEE Transactions on Instrumentation and Measurement 74: Article No. 1002511. https://doi.org/10.1109/TIM.2025.3533664

[30] LiuY, WangY, WangH, DouS, TianH, GanW, YuanQ. Dynamic template directed construction of three-dimensional porous bismuth aerogels for high-rate Na-ion storage. Journal of Materials Chemistry A, 2023, 1111: 5945-5955

[31] LiuZ, GuanX, ZhangL, ZhangY, PeiY, LiuC, ZhouH, LiuB, YanK. Investigations of dynamics of a single spark-induced bubble in saline water. Journal of Physics D-Applied Physicsv, 2021, 547075203

[32] LuM, ZhouB, BuZ. Two-stage distillation-aware compressed models for traffic classification. IEEE Internet of Things Journal, 2023, 1016: 14152-14166

[33] LuZ, ChenY, LiS, MaM. SDGC-YOLOv5: A more accurate model for small object detection. Artificial Neural Networks and Machine Learning, ICANN 2023, PT VII, 2023, 14260: 199-209

[34] MangS, YuY, ZhangQ. Numerical models for interactions of ablating-capillary plasma with air and bulk-loaded liquid. IEEE Transactions on Plasma Science, 2018, 468: 3065-3074

[35] PaquetteY, PelloneC, FivelM, GhigliottiG, JohnsenE, FrancJP. Modeling fluid-structure interaction in cavitation erosion: preliminary results. Journal of Physics: Conference Series, 2015, 9thInternationalSymposiumonCavitationCAV2015656012053

[36] PietronM, ZurekD, SniezynskiB. Speedup deep learning models on GPU by taking advantage of efficient unstructured pruning and bit-width reduction. Journal of Computational Science, 2023, 67101971

[37] RagunathanS. Numerical simulation of the flow field of a confined, submerged slot jet impinging on an oscillating surface: a parametric study. Proceedings of the ASME Fluids Engineering Division Summer Meeting, West Conshohocken: ASTM 1A: Paper No: FEDSM2016-7743, 2016

[38] RozhdestvenskyKV. Dynamics of vapor bubble in a variable pressure field. Journal of Marine Science and Application, 2022, 213: 83-98

[39] RuanX, LiuY, YuanC, LiB, HuW, LiY, MaybankS. EDP: An efficient decomposition and pruning scheme for convolutional neural network compression. IEEE Transactions on Neural Networks and Learning Systems, 2021, 3210: 4499-4513

[40] SchontagP, NakathD, RoehrlS, KoeserK. Towards cross domain transfer learning for underwater correspondence search. Image Analysis and Processing, 2022, ICIAP2022PTIIICham: Springer, 13233: 461-472

[41] ThanhHV, ThuNT, PhuongTM, VuongND, JoKHIeee. Inverted bottleneck convolution module for Yolov8. 2024 33RD International Symposium on Industrial Electronics, 2024, Ulsan. ISIE 1-6.

[42] TonellottoN, MacdonaldC, OunisI. Efficient query processing for scalable web search. Foundations and Trends in Information Retrieval, 2018, 124-5: 319-500

[43] TrivediC. A review on fluid structure interaction in hydraulic turbines: A focus on hydrodynamic damping. Engineering Failure Analysis, 2017, 77: 1-22

[44] WangX, HuangF, ChengY. Computational performance optimization of support vector machine based on support vectors. Neurocomputing, 2016, 211: 66-71

[45] XianY, YuY, LianY, FanJ, WangZ. An EA-based pruning on improved YOLOv3 for rapid copper elbow surface defect detection. Engineering Applications of Artificial Intelligence, 2023, 123PartC106412

[46] XiaoY, XiangH, WangT, WangY. Enhanced Industrial Action Recognition Through Self-Supervised Visual Transformers. IEEE Access, 2024, 12: 134133-134143

[47] XieY, ChenT, XuJAdvancing underwater acoustic target recognition via adaptive data pruning and smoothness-inducing regularization, 2023, 230411907

[48] ZhangAM, LiSM, XuRZ, PeiSC, LiS, LiuYL. A theoretical model for compressible bubble dynamics considering phase transition and migration. Journal of Fluid Mechanics, 2024, 999A58

[49] ZhangW, AnY. Instability of a bubble chain. Physical Review E, 2013, 875053023

[50] ZhaoGL, XuZY. Coastal vessel target detection model based on improved YOLOv7. Journal of Marine Science and Application, 2025

[51] ZhengLN, XuanP, HuangJ, LiJL. Development and application of spark-induced breakdown spectroscopy. Spectroscopy and Spectral Analysis, 2023, 433: 665-673

[52] ZhouJ, YangT, ZhangW. Underwater vision enhancement technologies: a comprehensive review, challenges, and recent trends. Applied Intelligence, 2023, 533: 3594-3621

[53] Zhou J, Zhang B, Zhang D, Vivone G, Jiang Q (2024) DTKD-Net: Dual-teacher knowledge distillation lightweight network for water-related optics image enhancement. IEEE Transactions on Geoscience and Remote Sensing 62: Article No. 4207213. https://doi.org/10.1109/TGRS.2024.3422667

RIGHTS & PERMISSIONS

Harbin Engineering University and Springer-Verlag GmbH Germany, part of Springer Nature