摘要
海洋工程的日常任务经常会利用水下航行器来探测水下环境,针对水下复杂环境采集的图像目标容易出现局部特征信息丢失,导致漏检、检测精度低的问题,提出基于增强局部特征的水下目标检测方法。在主干网络采用Faster Block增强图像特征提取能力;利用归一化注意力模块抑制不显著的特征信息来提高网络的效率;构建集中特征增强金字塔池化模块增强对水下复杂背景下目标局部特征信息的捕获能力;改进损失函数提高网络模型对水下图像目标预测效果。实验结果表明,该方法平均精度相较于原模型提升了1.5百分点,网络推理速度为36.4,能够有效地提升水下目标的检测精度。
In daily tasks related to ocean engineering,underwater vehicles are often employed to explore the underwater environment.Addressing the issue where image targets collected in complex underwater environments are prone to losing local feature information,which subsequently leads to missed detections and reduced detection accuracy,an underwater target detection method based on enhanced local features is proposed.The backbone network incorporates Faster Block to enhance image feature extraction capabilities.A normalized attention module is utilized to suppress non-significant feature information,thereby improving network efficiency.Additionally,a centralized feature enhancement pyramid pooling module is constructed to bolster the ability to capture local feature information in complex underwater backgrounds.By improving the loss function,the prediction accuracy of the network model for underwater image targets is enhanced.Experimental results demonstrate that this method improves the average accuracy by 1.5 percentage points compared to the original model,with a network inference speed of 36.4 frames per second,effectively enhancing the detection accuracy of underwater targets.
作者
张银胜
陈戈
张培琰
童俊毅
单梦姣
单慧琳
ZHANG Yinsheng;CHEN Ge;ZHANG Peiyan;TONG Junyi;SHAN Mengjiao;SHAN Huilin(School of Electronic&Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Electronic&Information Engineering,Wuxi University,Wuxi 214105,China)
出处
《中国测试》
北大核心
2025年第1期151-158,共8页
China Measurement & Test
基金
国家自然科学基金(62071240,62106111)
无锡市“太湖之光”科技攻关(基础研究)项目(K20241047)
无锡学院2023年教改研究课题(XYJG2023010,XYJG2023011)。
关键词
深度学习
目标检测
水下图像
部分卷积
注意力机制
局部特征
deep learning
target detection
underwater images
partial convolution
attention mechanism
local feature
作者简介
张银胜(1975-),男,江苏泰州市人,副教授,硕士研究生导师,博士,主要从事图像处理、人工智能的研究。
