计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (23): 211-218.DOI: 10.3778/j.issn.1002-8331.2207-0001

• 图形图像处理 • 上一篇    下一篇

结合轻量级特征提取网络的舰船目标检测算法

李登峰,高明,叶文韬   

  1. 西安工业大学 光电工程学院,西安 710021
  • 出版日期:2023-12-01 发布日期:2023-12-01

Ship Target Detection Algorithm Combined with Lightweight Feature Extraction Network

LI Dengfeng, GAO Ming, YE Wentao   

  1. College of Optoelectronic Engineering, Xi’an Technological University, Xi’an  710021, China
  • Online:2023-12-01 Published:2023-12-01

摘要: 由于海面舰船目标的检测设备受算力和存储空间所限,海面目标难以识别,所以舰船目标检测模型的大小需要轻量,且保持较高检测精度。针对如何平衡模型的大小和检测精度的问题,提出了基于轻量级特征提取网络的舰船目标检测算法。设计轻量级特征提取网络Ghost_ECA,在Ghostbottleneck中融入轻量级注意力机制ECA,模型轻量化的同时提高特征提取能力;用12个桶形排列的Ghost_ECA替换YOLOX的特征提取网络,降低模型的计算量和参数量;更换优化器为Adam,加快了模型的收敛速度;引入Focal loss损失函数,减少正负数据集样本不平衡问题,提高了检测精度。通过舰船数据集的对比实验表明:改进后的YOLOX模型相比于原YOLOX模型,模型的计算量和参数量分别降低27.54%和36.57%,检测精度高达89.86%。

关键词: 图像处理, 目标检测, YOLOX, 注意力机制, 轻量化网络

Abstract: Because the detection equipment of sea surface ship target is limited by computing power and storage space, it is difficult to identify sea surface target, so the size of ship target detection model needs to be lightweight and maintain high detection accuracy. Aiming at the problem of how to balance the size and detection accuracy of the model, a ship target detection algorithm based on lightweight feature extraction network is proposed. Design of lightweight feature extraction network Ghost_ECA, it integrates the lightweight attention mechanism ECA into Ghostbottleneck to lighten the model and improve the ability of feature extraction.The feature extraction network of YOLOX is replaced by 12 barrel arranged Ghost_ECA to reduce the amount of calculation and parameters of the model. Changing the optimizer to Adam accelerates the convergence speed of the model.The Focal loss function is introduced to reduce the sample imbalance of positive and negative data sets and improve the detection accuracy. The comparative experiments of ship data sets show that compared with the original YOLOX model, the amount of calculation and parameters of the improved YOLOX model are reduced by 27.54% and 36.57% respectively, and the detection accuracy is as high as 89.86%.

Key words: image processing, object detection, YOLOX, attention mechanism, lightweight network