Neural Network Used to Identify House Numbers in Google Street View
January 16, 2014
By Ed Silverstein, TMCnet Contributor
Google (News - Alert) Street View–which is used in Google Maps and Google Earth–provides views of many streets found around the world. Initially launched in 2007 in a select number of U.S. cities, it has since expanded to many locations worldwide.
Recently, new information has been released on how it identifies house numbers. It uses something called a “neural network.” It is best understood as a “computing network modeled on animal nervous systems,” according to a recent article appearing in Gizmodo. The neural network is used to identify house numbers from image data.
As part of the process, Google searches images to find and read the house numbers. Then, the information helps to locate a building in the database. It was also revealed that there are 11 levels of neurons which are used to locate house numbers in images. The process is very efficient and speedy.
"We can, for example, transcribe all the views we have of street numbers in France in less than an hour using our Google infrastructure," engineers explained in a recent Arxiv essay.
In the essay, the engineers also note the success of their overall approach. “We evaluate this approach on the publicly available SVHN [Street View House Numbers] dataset and achieve over 96 percent accuracy in recognizing complete street numbers,” they said. “We show that on a per-digit recognition task, we improve upon the state-of-the-art and achieve 97.84 percent accuracy. We also evaluate this approach on an even more challenging dataset generated from Street View imagery containing several tens of millions of street number annotations and achieve over 90 percent accuracy. Our evaluations further indicate that at specific operating thresholds, the performance of the proposed system is comparable to that of human operators. To date, our system has helped us extract close to 100 million physical street numbers from Street View imagery worldwide.”
The process aims to do as well as human operators who would be able to accurately spot numbers 98 percent of the time, according to an article from MIT Technology Review. That is impressive, but still not perfect. “That two percent of misidentified numbers is still a thorn in the team’s side,” the article said.
In analyzing their results, the tech review article questions if a similar method would be able to find telephone numbers on business signs or number plates. But street numbers are five numerals or less – with these other kinds of numbers generally being longer, and may not work as well.
There are other possible uses of the method. Using a single neural network could be used in general text transcription or in speech recognition, the engineers said. That could lead to some commercial applications that may find demand in the current and future marketplace.
In the meantime, the method used by Google deserves praise as one of its many success stories. “Google can rest assured that it has made a significant step forward in character extraction and recognition: the localization and identification of numbers by a single neural network,” according to the MIT (News - Alert) Technology Review article.
Edited by Ryan Sartor