车牌定位/检测公共数据集 [英] Public data set for licence plate localization/detection
问题描述
我一直在阅读了很多关于车牌定位/检测算法的论文，如这里所说的：
<一个href="http://stackoverflow.com/questions/4707607/what-are-good-algorithms-for-vehicle-license-plate-detection">What是好的算法的车牌检测？
我希望能对汽车的图像与已知的牌照号码的数据集上运行我的，所以我可以测量误差。
闲逛了相当数量后，我找不到任何此类公开数据集。有谁知道对否？如果不是这样，我的下一步将通过电子邮件发送文件的作者。
我不知道它是如何非法发表在公共场所采取任何形式的图像测试数据集既然已经做了，他们有被用在论文对交叉检查结果的目的，但我不知道什么级别的权限越来越有必要发布的数据集。它使你要想想你走之前，并采取了一堆照片，然后把它们发布。我知道，谷歌的街景模糊了所有的板，彻底搞乱了什么应该是一个几乎无限的金矿板块分析数据。它仍然是一个金矿自然场景数据小车定位的，但令人沮丧的从板块的角度来看。这或许可以解释为什么它是如此难以找到什么好的数据集一应俱全。你可能会认为我们将有板数以万计的一个相当标准的集合的地方。一位谷歌图片搜索车回实际上产生更少的结果比预期的要，你必须做一些采摘樱桃淘汰图像不工作。
在媒体实验室网站发布的@Abosamra真的很好，虽然他们是希腊，而不是美国的板块。
使用 http://www.vision.caltech.edu/html-files /archive.html 中的汽车1999年（后）2图像是巨大的，但也有只有其中的126。这是迄今为止美国最好的数据集是立即可下载，我发现，其中包括轿车在自然场景与板清晰易读。
这同一站点特征大量重复的2001的图像和多分辨率太低是用于测试的OCR，尽管它们可以被用于测试板定位有用和有超过500人
牛津大学是由伴随着1155的图像车太小，OCR，但可能是汽车检测过程中有用的一个更大的数据集的第一个链接重新发布的加州理工学院车
http://www.robots.ox.ac.uk/~ VGG / data3.html
麻省理工具有在第一次出现有为数据集但随后事实证明，在板太小做OCR，并且以ppm格式，所以可能需要额外的一个额外的步骤，以将它们转换为更通常使用的文件格式。他们可能是本地化测试是有用的。
http://cbcl.mit.edu/projects/cbcl/软件数据集/ CarData1Readme.html
加州大学圣地亚哥分校有一组约878图像，但需要从LOUKA Dlagnekov或塞尔Belongie获得权限访问该数据集。 http://vision.ucsd.edu/belongie-grp/research/carRec /car_data.html
如果你不介意使用在美国以外的板块，你可以找到更多的数据集。这里是一个，例如是在克罗地亚，但具有510的图像，这是良好的质量。他们有点放大，因此不太适合测试，以拉车出自然场景的能力，但非常好识别板。
http://www.zemris.fer.hr/projects/ LicensePlates /（简体）/ rezultati.shtml
由于真正的板块都很难找到汽车的场面，有可能使使用伪造板的数据集。这里是一个网站，可以制造假板块：
的假板可能是理想的用于制造板的测试数据集本身的分割，以及问题的，有很多不同的背景板和字体字符识别阶段
所以，唯一剩下的东西还真是找到一种方法来嵌入一些随机生成的假板与板模糊一些图像。
我会更新我的职务，如果我找到更多的。
I've been reading a lot of papers about licence plate localization/detection algorithms, such as those mentioned here:
What are good algorithms for vehicle license plate detection?
I was hoping to run mine on a data set of car images with known licence numbers so I can measure the error.
After a fair amount of poking around I can't find any publicly available data sets of this kind. Does anyone know of any? If not, my next step will be emailing paper authors.
I'm not sure how illegal it is to have published test datasets of images of any kind taken in a public space, since it has been done and they have been used in papers for the purpose of cross-checking results, but I don't know what level of permission getting was necessary to publish the dataset. It does make you want to think before you go out and take a bunch of pictures and then publish them. I know that Google's Street view blurs out all the plates, thoroughly messing up what should be a nearly unlimited gold-mine of plate analysis data. It's still a goldmine of car localization from natural scene data, but frustrating from a plate perspective. That might explain why it is so difficult to find any good datasets readily available. You would think that we would have a fairly standard collection of tens of thousands of plates somewhere. A Google image search for "back of car" actually produces fewer results than would be expected and you have to do some cherry picking to weed out the images that don't work.
The Medialab Website posted by @Abosamra is really good, though they are Greek, not American plates.
With http://www.vision.caltech.edu/html-files/archive.html the Cars 1999 (Rear) 2 images are great, though there are only 126 of them. This is by far the best American dataset that is instantly downloadable that I have found, including cars in natural scenes with plates clearly readable.
The 2001 images from the same site feature lots of repeats and are mostly too low resolution to be useful for testing OCR, though they could be used for testing plate localization and there are over 500 of them.
Oxford is re-publishing the Caltech cars from the first link along with a much larger dataset of 1155 images of cars too small for OCR but might be useful for the car detection process:
http://www.robots.ox.ac.uk/~vgg/data3.html
MIT has a dataset which at first appears promising but then it turns out that the plates are too small to do OCR, and are in ppm format, so may require an extra an extra step to convert them to a more generally used file format. They could be useful for localization tests.
http://cbcl.mit.edu/projects/cbcl/software-datasets/CarData1Readme.html
UCSD has a set of about 878 images but requires getting permission from Louka Dlagnekov or Serge Belongie to access the dataset. http://vision.ucsd.edu/belongie-grp/research/carRec/car_data.html
If you don't mind using plates outside the US, you can find more datasets. Here is one, for example that is in Croatia, but has 510 images, which are good quality. They are a bit zoomed in, so not great for testing the ability to pull the car out of the natural scene, but very good for identifying plates.
http://www.zemris.fer.hr/projects/LicensePlates/hrvatski/rezultati.shtml
Since real plates are so hard to find in car scenes, it might be possible to make a dataset using fake plates. Here is a site that can create fake plates:
The fake plates could be ideal for making a test dataset of plates themselves for the segmentation, and character recognition phases of the problem with lots of different plate backgrounds and fonts.
So the only thing left really is to find a way to embed some randomly created fake plates into some images with blurred plates.
I may update my post if I find more.
这篇关于车牌定位/检测公共数据集的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！
