本文为芬兰阿尔托大学(作者:OlaviStenroos)的硕士论文,共75页。
目标检测是当前基于机器学习的计算机视觉的一个子领域。在过去的数十年中,机器学习领域被所谓的深度神经网络所主导,其基础是基于计算能力和数据可用性的提高。卷积神经网络(CNN)是神经网络的一个子类,非常适合于处理与图像相关的任务。神经网络通过训练在图像中寻找不同的特征,例如边缘、角和颜色差异,并将它们组合成更复杂的形状。对于目标检测,系统必须同时估计可能目标的位置并对它们进行分类。
对于本硕士论文,我们回顾了当前关于卷积目标检测的文献,并测试了其中一种方法的可实现性。我们发现,尽管卷积目标检测已经优于其它目标检测方法,但卷积目标检测作为一项技术仍在不断发展。由于数据集和预训练网络的免费可用性,可以在不访问专用硬件的情况下实现深度神经网络。预训练网络也可以用作训练新网络的起点,从而减少昂贵的训练时间。
在实验部分,我们使用MATLAB和MatConvNet实现了快速R-CNN,并在两个不同流量的数据集上测试了一个通用的目标检测器。我们发现,与原来的卷积目标检测方法相比,快速R-CNN更加准确快速,并且能够运行在个人计算机上。诸如更快速R-CNN和SSD的高级算法,可以进一步提高快速CNN的处理速度。我们还实验了一个基于几何的场景估计模型,据报道该模型能够提高目标检测方法的精度。而实际上,我们从研究中发现这种模型并没有带来改进,因此将整个场景建模与卷积网络相结合是一个需要进一步研究的潜在课题。
Object detection is a subfield of computervision that is currently heavily based on machine learning. For the pastdecade, the field of machine learning has been dominated by so-called deepneural networks, which take advantage of improvements in computing power anddata availability. A subtype of a neural network called a convolutional neuralnetwork (CNN) is well-suited for image-related tasks. The network is trained tolook for different features, such as edges, corners and colour differences,across the image and to combine these into more complex shapes. For objectdetection, the system has to both estimate the locations of probable objectsand to classify these. For this master’s thesis, we reviewed the currentliterature on convolutional object detection and tested the implementability ofone of the methods. We found that convolutional object detection is stillevolving as a technology, despite outranking other object detection methods. Byvirtue of free availability of datasets and pretrained networks, it is possibleto create a functional implementation of a deep neural network without accessto specialist hardware. Pretrained networks can also be used as a startingpoint for training new networks, decreasing costly training time. For theexperimental part, we implemented Fast R-CNN using MATLAB and MatConvNet andtested a general object detector on two different traffic-related datasets. Wefound that Fast R-CNN is relatively precise and considerably faster than theoriginal convolutional object detection method, R-CNN, and can be implementedon a home computer. Advanced methods, such as Faster R-CNN and SSD, improve thespeed of Fast R-CNN. We also experimented with a geometrybased scene estimationmodel, which was reported to improve the precision of a previous generationobject detection method. We found that with our implementation of Fast R-CNNthere was no such improvement, although further adjustments are possible.Combining whole scene modelling with convolutional networks is a potential subjectof further study.
1 引言
2 项目背景
3 卷积目标检测
4 实验设计
5 具体实现
6 性能评估
7 讨论
8 结论
更多精彩文章请关注微信号: