大工至善|大学至真分享 http://blog.sciencenet.cn/u/lcj2212916

博文

[转载]【计算机科学】【2016.11】利用卷积神经网络识别车辆姿态和类别

已有 1474 次阅读 2020-6-4 17:33 |系统分类:科研笔记|文章来源:转载

本文为德国达姆施塔特工业大学(作者:Christoph Münker)的硕士论文,共99页。

 

卷积神经网络(ConvNets)是一种非常适合于图像处理的专用神经网络,近年来在许多机器学习问题上取得了巨大的成功。因此,我们感兴趣的是ConvNets在车辆图像上用于车辆姿态和类别分类的性能。为了将ConvNets应用于这两个任务,我们开发了自己的ConvNet架构,并将其与三种最先进的ConvNet架构进行了比较。此外,我们还研究了这两个任务之间的协作能力,因为车辆分类取决于车辆的姿态。因此,我们研究了两种方法。第一种方法是多任务学习,在同一模型上同时学习两个任务。作为第二种方法,我们使用专家学习者。它利用图像上检测到的姿态来选择一个经过特殊训练的模型来进行车辆分类。作为数据基础,使用了Yang等人的“综合汽车”数据集。从这个数据集中,我们使用97269张图像和967辆不同的车辆来训练和测试ConvNet。对车辆姿态的标签进行了手动扩展和改进。对于车辆类,我们定义了一个新的分类,并且可以半自动地分配标签。最后用八个方向描述车辆姿态,用八个类别描述车辆类别。结果表明,车辆姿态的学习不是问题。具有默认设置的单任务模型的错误率为1.12%车辆分类的效果不太好,默认设置的错误率为28.89%。但多任务学习提高了车辆分类的学习性能,使我们的最佳单任务模型的错误率达到26.04%。集成学习将错误率减少到24.87%

 

Convolution Neural Networks (ConvNets) arespecialized Neural Networks, which are well suited for the processing of imagesand had enormously success on many machine learning problems in the last years.Therefore, we are interested in the performance of ConvNets for theclassification of vehicle pose and vehicle class on images of vehicles. Toapply ConvNets on these two tasks, we have developed our own ConvNetarchitecture, and compared this to three state-of-the-art ConvNetarchitectures. Furthermore, we study the capability of the cooperation betweenthese two tasks, because the appearance of the vehicle class is dependent onthe pose of the vehicle. Therefore, we investigate two approaches. The firstapproach is the multi-task learning, which simultaneously learns the two taskson the same model. As the second approach, we use an expert learner. It usesthe detected pose on the image to chose a specially trained model on this posefor the classification of the vehicle class. As data basis is used “TheComprehensive Cars” dataset from Yang et al. [57]. From this dataset, we use97,269 images with 967 different vehicles to train and test the ConvNet. Thelabels for the vehicle pose are manually expanded and improved. For the vehicleclass, we define a new categorization and could semi-automatically assign thelabels. At the end the vehicle pose is described by eight orientations and the vehicleclass by eight classes. The results show, that the learning of the vehicle poseis not a problem. The single-task model with default settings get an error rateof 1.12%. The vehicle class is not so good and get an error rate of 28.89% withthe default settings. But the multi-task learning increases the performance ofthe vehicle class, so that we reach an error rate of 26.04% with our bestsingle model. The ensemble learning improves the error to 24.87%.

 

1. 引言

2. 机器学习、神经网络与卷积神经网络基础

3. 相关工作

4. 数据集

5. 图像预处理与分割

6. 研究方法

7. 评估

8. 结论与未来工作展望



更多精彩文章请关注公众号:205328s611i1aqxbbgxv19.jpg




https://blog.sciencenet.cn/blog-69686-1236444.html

上一篇:[转载]【信息技术】【2009.04】人体目标的检测与跟踪
下一篇:[转载]【统计学】【2016】【含源码】基于时间序列分析的度量数据告警预测
收藏 IP: 60.169.68.*| 热度|

0

该博文允许注册用户评论 请点击登录 评论 (0 个评论)

数据加载中...

Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2024-3-29 21:45

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社

返回顶部