基于改进YOLOv8算法的复杂场景船舶目标检测

doi:10.12068/j.issn.1005-3026.2025.20240118

摘要/Abstract

摘要：

为了提高复杂场景下船舶目标检测的精度和鲁棒性，基于YOLOv8算法进行模型改进.在骨干层引入CD3模块并嵌入无参注意力SimAM模块；在颈部层引入注意力尺度序列融合ASF（attention-based scale sequence fusion）模块；在头部层增加了检测头用于预测输出；采用剪枝降低模型计算量；最后使用蒸馏进一步提升模型性能.实验在阿里天池官网提供的复杂场景船舶检测数据集上进行验证，结果表明，改进后模型的AP₅₀和AP相比于YOLOv8分别提升4.7%和2.9%，召回率和精确率分别提升3.2%和4.2%，而模型参数量减少56.1%，计算量降低30.5%.改进后的模型在减少参数量的同时，提升了总体性能.

关键词: YOLOv8算法, CD3, ASF-YOLO, SimAM, 剪枝蒸馏

Abstract:

To improve the accuracy and robustness of ship target detection in complex scenarios， modifications were implemented based on the YOLOv8 algorithm. The CD3 module was introduced in the backbone layer with the parameter-free attention SimAM module embedded. The attention-based scale sequence fusion （ASF） module was incorporated in the neck layer， and an additional detection head was added to the head layer for prediction output. Pruning was adopted to reduce the computations of the model， followed by distillation to further improve model performance. The experiment was conducted on the complex scenario ship detection dataset from Alibaba Tianchi for verification. The results demonstrate that compared with YOLOv8， the improved model achieves increases of 4.7% in AP₅₀ and 2.9% in AP， respectively. Recall and precision are improved by 3.2% and 4.2%， while model parameters and computations are reduced by 56.1% and 30.5%. The optimized model thus improves overall performance while reducing parameters.

Key words: YOLOv8 algorithm, CD3, ASF-YOLO, SimAM, pruning and distillation

中图分类号:

TP 391.4

车晓辰, 马淑华, 郭泽旭, 沙晓鹏. 基于改进YOLOv8算法的复杂场景船舶目标检测[J]. 东北大学学报（自然科学版）, 2025, 46(12): 38-47.

Xiao-chen CHE, Shu-hua MA, Ze-xu GUO, Xiao-peng SHA. Ship Target Detection in Complex Scenarios Based on Improved YOLOv8 Algorithm[J]. Journal of Northeastern University(Natural Science), 2025, 46(12): 38-47.

图/表 8

图1 改进的YOLOv8网络结构

Fig.1 Improved YOLOv8 network architecture

图2 TFE结构图CBS表示一个由卷积层（convolution）、批归一化层（batch normalization）和SiLU激活函数组成的序列模块.

Fig.2 TFE structure diagram

图3 剪枝通道对比

Fig.3 Comparison of pruning channels

图4 知识蒸馏结构图

Fig.4 Structural diagram of knowledge distillation

表1 消融实验结果

Table 1 Results of ablation experiment

模型	基准模型及其改进	AP₅₀	AP	精确率	召回率	计算量/GFLOPs	参数量×10^-6
A	YOLOv8	0.575 14	0.335 09	0.682 93	0.509 46	8.2	3.011 1
B	A+CD3	0.582 01	0.337 91	0.661 61	0.517 03	7.4	2.668 9
C	B+(ASF+检测头)	0.603 01	0.349 79	0.685 91	0.538 99	11.4	2.234 9
D	C+SimAM	0.609 97	0.362 16	0.720 73	0.523 11	11.4	2.234 9
E	D+剪枝	0.604 83	0.349 49	0.708 05	0.508 71	5.7	1.320 9
F	E+蒸馏	0.622 36	0.363 86	0.725 06	0.541 26	5.7	1.320 9

图5 模型A的部分检测结果（a）—轻微干扰；（b）—非船舶轻微遮挡；（c）—远距离、小目标；（d）—同类型船舶混淆干扰且遮挡面积较大；（e）—背景复杂、干扰较多；（f）—远距离且船舶密集排列.

Fig.5 Partial detection results of model A

图6 模型F的部分检测结果（a）—轻微干扰；（b）—非船舶轻微遮挡；（c）—远距离、小目标；（d）—同类型船舶混淆干扰且遮挡面积较大；（e）—背景复杂、干扰较多；（f）—远距离且船舶密集排列.

Fig.6 Partial detection results of model F

表2 与其他SOTA检测器对比实验结果

Table 2 Comparison of experimental results with other SOTA detectors

模型	AP₅₀	AP	精确率	召回率	计算量/GFLOPs	参数量×10^-6
YOLOv5-n	0.565	0.324	0.663	0.504	7.2	2.51
YOLOv5-m	0.592	0.342	0.669	0.533	24.0	9.12
YOLOv6-n	0.577	0.337	0.708	0.481	11.9	4.24
YOLOv7-tiny	0.605	0.324	0.595	0.583	13.2	6.01
gold_yolo-n	0.587	0.335	0.699	0.518	12.1	5.60
ResNet50-FasterRCNN	0.548	0.298	0.543	0.387	251.4	41.35
SSD	0.521	0.281	0.643	0.558	122.3	48.31
CCAT-YOLOv8n	0.581	0.338	0.651	0.518	15.9	5.72
YOLOv8n-Bi-FPN-EMA	0.587	0.336	0.687	0.528	7.8	2.52
模型F(本文模型)	0.622	0.364	0.725	0.541	5.7	1.32

参考文献 22

[1]	鄢然，王帅安，周煜圣. 区块链技术在航运业的应用综述［J］. 交通运输工程与信息学报， 2022， 20（3）： 1-14.
	Yan Ran， Wang Shuai-an， Zhou Yu-sheng. Application of blockchain technology in the shipping industry［J］. Journal of Transportation Engineering and Information， 2022， 20（3）：1-14.
[2]	周旗开，张伟，李东锦，等. 基于改进YOLOv5s的光学遥感图像舰船分类检测方法［J］. 激光与光电子学进展， 2022， 59（16）： 1628008.
	Zhou Qi-kai， Zhang Wei， Li Dong-jin， et al. Ship detection and classification method for optical remote sensing images based on improved YOLOv5s［J］. Laser& Optoelectronics Progress， 2022， 59（16）： 1628008.
[3]	Li Y D， Zhang S S， Wang W Q. A lightweight faster R-CNN for ship detection in SAR images［J］. IEEE Geoscience and Remote Sensing Letters， 2022， 19： 4006105.
[4]	Jiang Z K， Su L， Sun Y X. YOLOv7-ship： a lightweight algorithm for ship object detection in complex marine environments［J］. Journal of Marine Science and Engineering， 2024， 12（1）： 179-190.
[5]	Zhao Q C， Wu Y Q， Yuan Y B. Ship target detection in optical remote sensing images based on E2YOLOX-VFL［J］. Remote Sensing， 2024， 16（2）： 318-340.
[6]	Lyu Z L， Wang C Y， Sun X J， et al. Real-time ship detection system for wave glider based on YOLOv5s-lite-CBAM model［J］. Applied Ocean Research， 2024， 144： 103833.
[7]	Zhao K， Liu R T， Wang S Y， et al. ST-YOLOA： a Swin-Transformer-based YOLO model with an attention mechanism for SAR ship detection under complex background［J］. Frontiers in Neurorobotics， 2023， 17： 1170163.
[8]	Jocher G， Chaurasia A， Qiu J. Ultralytics YOLO （version 8.1.9）［EB/OL］. （2023-01-10）［2024-02-24］. .
[9]	Wang W H， Dai J F， Chen Z， et al. InternImage： exploring large-scale vision foundation models with deformable convolutions［C］//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver： IEEE， 2023： 14408-14419.
[10]	Kang M， Ting C M， Ting F F， et al. ASF-YOLO： a novel YOLO model with attentional scale sequence fusion for cell instance segmentation［J］. Image and Vision Computing， 2024， 147： 105057-105096.
[11]	Yang L X， Zhang R Y， Li L D， et al. SimAM： a simple， parameter-free attention module for convolutional neural networks［C］//International Conference on Machine Learning. Online： PMLR， 2021： 11863-11874.
[12]	Zhu X Z， Hu H， Lin S， et al. Deformable ConvNetsv2： more deformable， better results［C］//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition（CVPR）. Long Beach： IEEE， 2019： 9300-9308.
[13]	Lee J， Park S， Mo S， et al. Layer-adaptive sparsity for the magnitude-based pruning［EB/OL］. （2021-05-09）［2024-02-26］. .
[14]	Li H， Kadav A， Durdanovic I， et al. Pruning filters for efficient ConvNets［EB/OL］. （2017-03-17）［2024-02-26］. .
[15]	Howard A， Sandler M， Chen B， et al. Searching for MobileNetV3［C］// 2019 IEEE/CVF International Conference on Computer Vision （ICCV）. Seoul： IEEE， 2019： 1314-1324.
[16]	Blalock D， Ortiz J J G， Frankle J， et al. What is the state of neural network pruning？［EB/OL］. （2020-03-06）［2024-02-26］. .
[17]	Molchanov P， Tyree S， Karras T， et al. Pruning convolutional neural networks for resource efficient inference［EB/OL］. （2017-06-08）［2024-02-26］. .
[18]	Howard A G， Zhu M L， Chen B， et al. MobileNets： efficient convolutional neural networks for mobile vision applications［EB/OL］. （2017-04-17）［2024-02-26］. .
[19]	Wang Q， Liu L， Yu W X， et al. BCKD： block-correlation knowledge distillation［C］//2023 IEEE International Conference on Image Processing （ICIP）. Kuala Lumpur： IEEE， 2023： 3225-3229.
[20]	Shu C Y， Liu Y F， Gao J F， et al. Channel-wise knowledge distillation for dense prediction［C］//2021 IEEE/CVF International Conference on Computer Vision （ICCV）. Montreal： IEEE， 2022： 5291-5300.
[21]	李斌，雷钧涵，郭毅. 基于CCA和Transformer的YOLOv8船舶目标检测算法［J］. 控制工程， 2024， 31（5）： 901-911.
	Li Bin， Lei Jun-han， Guo Yi. YOLOv8 ship target detection algorithm based on CCA and Transformer［J］. Control Engineering of China， 2024， 31（5）： 901-911.
[22]	惠卓凡，李鹏龙，沈烈，等. 基于改进YOLOv8的渔港船舶进出港目标检测与统计方法［J］. 大连海洋大学学报， 2024， 39（3）： 498-505.
	Hui Zhuo-fan， Li Peng-long， Shen Lie， et al. Detection and statistics method of ship entry and exit in a fishing port based on improved YOLOv8［J］. Journal of Dalian Ocean University， 2024， 39（3）：498-505.