Google makes for optimization of Google maps used OCR open source

Google has a his machine-learning project TensorFlow related technology text recognition in pictures open source made to. You interested is now in Github available . TensorFlow even has Google 2015 under the Apache license.

 TensorFlow (image: Google)

on the occasion of the publication of the Google employees Julian Ibarz type, software engineer in the brain team and Sujoy Banerjee, product manager of the ground truth team also insight into the use of the text recognition technology by the group. According to them, collect the street view vehicles of every day still millions of images. As a whole, now 80 billion high-resolution recordings exist for street view. In such numbers, it is simply impossible to evaluate the images manually.

one of the tasks of the Google ground truth team is therefore to develop methods for the automatic extraction of information from the images marked with geo-data and then use that to improve Google maps. Street names that are read from street signs are an important part of it . While multiple shots are used each, to improve the accuracy, to detect deviations in the notation and to normalize the different variants. That seems particularly difficult in France to be, that’s why the Google researchers illustrate the work with the software on the example of this country . The algorithm there to reach an accuracy of 84.2 percent and makes it far more powerful than earlier systems.

in addition, is he limited to street names, but let easily applied to the extraction of other information from street view imagery. An example is about the recognition of names of transactions based on their signboards.

Ibarz and Banerjee indicating that the focus of automatic text recognition (OCR) had traditionally been on scanned documents. The recognition of texts from shots “in the wild” but instead the researchers before all other tasks because there are partly covered or poorly readable texts, the recording angle ensures distortion or pictures can be blurred.

started with work to accomplish these tasks, automated Google 2008 occasion was at the time the call launched on Google initially not immediately on open ears to make obscure faces and license plates on the street-view shooting in some countries. After this was genomes in attack, man but apparently has recognized the other possibilities.

“we noticed that with sufficiently classified data could not only use machine learning to protect the privacy of our users, we could accumulate even Google maps automatically with relevant and current information”, explain the researchers. [Anmerkung der Redaktion: Eigentlich ging es hier nicht um die „Privatsphäre der Nutzer“, sondern um die Privatsphäre zufällig aufgenommener Unbeteiligter, die nicht unbedingt Google-Nutzer sein mussten].

 the system developed by Google can be in the image the company name the system developed by Google can be in the image the company name” Zelina pneus “correctly recognize, despite him no information about its location in the image. The model of tires sold by the company is not confused by the brand name visible in the image (image: Google)

one of the earlier results was the 2014 presented system for detection of house numbers. It was a crucial step to make Google maps more accurate, explain Ibarz and Banerjee. So far the accuracy was been improved so that over one-third of the collected addresses worldwide. In some countries, including Brazil, the percentage of the now more accurately mapped addresses lie even at 90 percent.

the technology was then based on a data bank with over a million street names from France on applied. In contrast to the house number recognition, it is necessary make sense to merge data from multiple images in the detection of street names may. Variable text (about road or street) and accessories (some dates for the House numbers) and abbreviations must also (about BGM.-Fritz Müller-Straße at the Mayor Fritz Müller-Straße) be recognized as such and consistently associated with the right road.

 processing unit (TPU), Google in its data centers using tensor, is content with less transistors per calculation than other processors. This results in a significantly higher performance per watt for machine learning, what results are ultimately more intelligent users in less time. (Bild: Google). processing unit (TPU), Google in its data centers using tensor, is content with less transistors per calculation than other processors. This results in a significantly higher performance per watt for machine learning, what results are ultimately more intelligent users in less time. (Image: Google).

the new system allow it along with the house number detection addresses in Google maps, where before either associated street names or house number were not known, to create directly from the images. “Now if a street view car on a newly built road, can our system analyze the tens of thousands of pictures, street names and house numbers extract and create correctly new addresses and geographically correct map”, so the researchers.

this was expanded to include detection of company names based on the facade of shops. The task here was to determine the name of the variety of information (chooses about manufacturers there products, notes on actions etc.). That succeed but now satisfactorily – partly because you once again have upgraded the computing power used in the background .

WEBINAR

what next – BB´s storage & co: the enterprise cloud!

get to know the building blocks of enterprise cloud platform in this audio Webinar. Learn how to achieve maximum freedom and flexibility for your applications. More outcome achieved with less input – specific application examples. [Update: the webinar has already occurred.] Register now and look at the record.

Be the first to comment

Leave a Reply