摘要
为了改善执法监管工作中人工查看视频方式的不便性,使用改进的DenseNet201模型依据是否悬挂国旗这一指标对船舶图像进行分类。在此过程中,以DenseNet模型为基础,首先在DenseNet模型网络结构中融入了注意力机制SENet模块,以提升船舶图像重要特征提取的指向性;其次使用亮度、对比度增强等方式扩充船舶图像数目并进行类别标注;最后使用ImageNet数据集下预训练的DensenNet201模型,使用迁移学习方式在预训练模型下使用船舶数据集进行参数微调,提升模型对于船舶图像的泛化能力。经试验验证,使用DenseNet模型以及改进DenseNet模型分别做图像分类,在验证集上准确率分别为85%和93%,提高了约8%的准确率,说明改进的DenseNet模型在船舶数据集上具有良好的分类性能。
To improve the inconvenience of manually viewing videos in law enforcement and supervision work,the improved DenseNet201 model is used to classify ship images based on whether the national flag is hoisted.In this process,based on the DenseNet model,the attention-mechanism SENet(Squeeze-and-Excitation Net)module is firstly integrated into the network structure of the DenseNet model to improve the directivity of important features extraction in ship images.Secondly,brightness and contrast enhancement are used to expand the number of ship images and label them.Finally,the DensenNet201 model pre-trained under the ImageNet data set is used.The transfer learning method is used to fine-tune the parameters of the ship data set under the pre-trained model,so as to improve the generalization ability of the model for ship images.The experimental results show that the DenseNet model and the improved DenseNet model are used for image classification respectively,and the accuracy of the validation set is 85%and 93%,which is about 8%higher,indicating that the improved DenseNet model has good classification performance on the ship data set.
作者
居云龙
周怡君
罗晨
孟波波
JU Yunlong;ZHOU Yijun;LUO Chen;Meng Bobo(School of Software,Southeast University,Suzhou 215000,Jiangsu,China;School of Mechanical Engineering,Southeast University,Nanjing 211189,China;Jiangsu Century Information Technology Co.,Ltd.,Nanjing 210012,China)
出处
《船舶标准化工程师》
2024年第5期80-86,共7页
Ship Standardization Engineer
