基于更快区域卷积神经网络的多视角船舶识别
Multiview ship recognition based on the faster region convolutional neural network
程静 1王荣杰 2曾光淼 1林安辉 2王亦春2
作者信息
- 1. 集美大学 轮机工程学院,福建 厦门 361021
- 2. 集美大学 轮机工程学院,福建 厦门 361021;福建省船舶与海洋工程重点实验室,福建 厦门 361021
- 折叠
摘要
针对在复杂海洋环境下采集船舶多视角图像难度大、不同视角下船舶外观差异显著的问题,本文以自制的不同类型的多艘船舶的多视角图像为数据集训练更快区域卷积神经网络模型,利用平均F1分数、平均精度和平均误检率作为评价指标评估更快区域卷积神经网络模型对不同视角船舶的识别性能,并通过识别不同船舶的F1分数和误检率分析更快区域卷积神经网络对不同质量、背景图像的识别能力.实验结果表明,更快区域卷积神经网络识别多角度船舶的平均F1分数为0.696 9,平均精度为92.88%,平均误检率为8.34%,即更快区域卷积神经网络对多视角船舶有较高的识别能力,但对于有雾或昏暗环境下的低像素图像识别能力明显下降.
Abstract
The appearance of ships varies greatly under different views,and the complex ocean environment increases the difficulty of obtaining multiview ship images.In this paper,the self-made multiview images of different types of ships were used as datasets to train the faster region convolutional neural network(faster R-CNN)model.Then,we evaluated the recognition performance of the faster R-CNN model for ships of different views by using the mean F1 score,mean accuracy,and mean log-average miss rate as evaluation indexes.Furthermore,the recognition ability of the faster R-CNN to different levels of quality and background images was analyzed by identifying the F1 scores and log-average miss rates of different ships.The experimental results showed that the mean F1 score,mean accuracy,and mean log-average miss rate of the faster R-CNN were 0.696 9,92.88%and 8.34%,respectively.The faster R-CNN has high recognition ability when processing multiview ship images;however,the same ability is significantly reduced when processing low-pixel images in foggy or dim environments.
关键词
多视角/船舶识别/视觉图像/更快区域卷积神经网络/目标检测/特征提取/深度学习/低分辨率图像Key words
multiview/ship recognition/vision image/faster region convolutional neural network(faster R-CNN)/object detection/feature extraction/deep learning/low resolution image引用本文复制引用
出版年
2023
国家科技期刊平台
NSTL
维普
万方数据