Banner
Computer

Google makes open source code used to optimize Google Maps

It is used to extract information from images. This includes street names or business names. The technology goes back to the work begun in 2008 to make faces and license plates unrecognizable in street-view recordings.

Google makes open source code used to optimize Google Maps

Google has made a technology for text recognition in its pictures of the TensorFlow machine project. It is now available to interested people at Github. TensorFlow itself has offered Google since 2015 under the Apache License.

On the occasion of the publication, the Google employees Julian Ibarz, Software Engineer in the Brain Team and Sujoy Banerjee, Product Manager of the Ground Truth Team also give insights into the use of the text recognition technology by the Group.According to them, the Street View vehicles still collect millions of images every day. In total, 80 billion high-resolution recordings are now available for Street View. With such figures, it is simply impossible to evaluate the recordings manually.

One of the tasks of the Google Ground Truth team is therefore to develop methods for the automatic extraction of information from the Geo-data, and then use it to improve Google Maps. An important part of these are road names, which are read off road signs. In doing so, a plurality of receptacles are used in each case in order to improve the accuracy, recognize deviations in the spelling, and normalize the different variants. This seems particularly difficult in France, so the Google researchers illustrate the work with the software of the example of this country. The algorithm achieves an accuracy of 84.2 percent there and is thereby considerably more powerful than previous systems.

Google Open Source Algorithm

In addition, it is not restricted to street names, but is easily applied to extracting other information from street-view images. One example is the recognition of the names of businesses on the basis of their company signs.

Ibarz and Banerjee point out that the focus of automatic text recognition (OCR) has traditionally been on scanned documents. The recognition of texts from recordings “in the wild” poses the researchers however quite different tasks, since texts are partly obscured or illegible, the recording angle for distortion or recordings can be blurred.

google open source technology

Google started to work on these tasks in an automated way. In 2008, Google’s first demand was not to open the door to faces-and-license plates in the street-view recordings in some countries. After this became in attacking genomes, however, one obviously recognized the other possibilities.

“We noticed that we could not only use machine data that was adequately classified to protect our users’ privacy, but would also automate Google Maps with relevant and up-to-date information,” the researchers explain. [Editor’s note: Actually, it was not the “privacy of the users”, but the privacy of randomly-taken uninvolved people who were not necessarily Google users].

Google Map Technology

One of the earlier findings was the 2014 home identification system. It was a crucial step to make Google Maps more accurate, Ibarz and Banerjee explain. Up to now, accuracy has been improved over more than one third of the registered addresses worldwide. In some countries, including Brazil, the percentage of the more accurate addresses is even 90 percent.

The technology was then applied to a database of over one million street names from France. In contrast to the house number recognition, it may be necessary to merge data from several images meaningfully in the recognition of street names. In addition, variable text (such as street or street) as well as additions (such as information on the house numbers) and abbreviations (for example, Bmi-Fritz-Müller-Strasse at the mayor Fritz-Müller-Strasse) must be recognized as such and assigned to the correct street uniformly become.

The new system, together with the house number recognition, allows to create addresses directly from the pictures in Google Maps, where previously the corresponding street name or the house number were not known. “When a street view car is driving on a newly built road, our system can analyze tens of thousands of images taken, extract street names and house numbers, and properly create new addresses and assign them geographically correctly,” the researchers said.

This has also been extended to the recognition of company names on the facade of store shops. The task here was to extract from the Vielz

Leave a Reply

Your email address will not be published. Required fields are marked *

Banner