基于改进FaceNet的河湖采砂船"船脸"识别算法
"Ship face"recognition algorithm for river and lake sand dredgers based on improved FaceNet
包学才 1陈豹 1吴灿锐 1汪忠喜 1占礼彬1
作者信息
- 1. 南昌工程学院信息工程学院,江西南昌 330099;南昌工程学院 江西省水信息协同感知与智能处理重点实验室,江西南昌 330099
- 折叠
摘要
为有效提升河湖采砂船智能化管理水平,提出了一种基于改进FaceNet的河湖采砂船"船脸"识别算法.首先在FaceNet算法网络的全局平均池化层后引入CA注意力模块,增强算法对于感兴趣区域的自适应关注能力;其次训练时在网络的最后引入线性层构建采砂船个体"船脸"识别器,将分类和识别的方法相结合共同应用于采砂船"船脸"识别;最后在训练时引入交叉熵损失函数,辅助原FaceNet算法中的三元组损失函数共同实现收敛.实验结果表明:改进的FaceNet算法对于白天场景下采砂船个体"船脸"目标识别的正确率比改进前提高了4.77%,达79.22%;夜间场景下目标识别的正确率提高了2.83%.研究成果适用于采砂船"船脸"识别任务,可为河湖采砂船的智能监管提供技术参考.
Abstract
In order to effectively improve the intelligent management level of river and lake sand dredgers,an improved FaceNet based"ship face"recognition algorithm for river and lake sand dredgers was proposed.Firstly,a CA attention module was intro-duced behind the global average pooling layer of the FaceNet algorithm network to enhance the adaptive attention ability for re-gions of interest.Secondly,a linear layer was introduced at the end of the network during training to construct an individual"ship face"recognizer for sand dredgers.The combination of classification and recognition methods was applied to the"ship face"rec-ognition of sand dredgers.Finally,Cross entropy Loss function was introduced into the training to assist the Triplet loss function in the original FaceNet algorithm to converge together.The experimental results showed that the accuracy of the improved FaceNet al-gorithm for identifying individual"ship face"objects on sand dredgers in daytime had increased by 4.77 percentage points com-pared tothat before the improvement,reaching79.22%.The accuracy of identifyingindividual"shipface"objects of sand dredg-ers in night had increased by 2.83 percentage points.This algorithm is suitable for the"ship face"recognition task of sand dredg-ers and can provide effective technical support for the intelligent supervision of river and lake sand dredgers.
关键词
采砂船监管/FaceNet/深度学习/卷积神经网络/目标识别Key words
supervision of sand dredgers/FaceNet/deep learning/convolutional neural network/object identification引用本文复制引用
基金项目
国家自然科学基金项目(61961026)
江西省科技厅重大科技研发专项"揭榜挂帅"制项目(20213AAG01012)
江西省水利厅科技项目(202223YBKT19)
出版年
2024
国家科技期刊平台
NSTL
万方数据