Google engineers told about the success in recognizing house numbers from StreetView photos
A team of Google engineers, led by a guy with a friendly surname Goodfellow ( Ian Goodfellow , graduated from CS University in 2009 and boasts a solid list of scientific publications), published on arxiv.org, in which she reported on her success in collecting information that the base has Google Street View project - in particular, on automatic recognition of house numbers with the same quality as a person would. An eloquent fact is that Google can get all the house numbers in France (more precisely, where the Google Mobile has traveled) in less than an hour.
The work is called Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks and it is about using neural networks for recognition purposes. As is obvious, the case is greatly complicated by many factors: the different locations of the house number, the color of the plate, its various inclinations, the quality of the plate itself, as well as its photographs, and a number of others.
As a result, the task was shifted to the modification of the DistBelief neural network with 11 levels of neurons (this is the same network that Google uses for “deep learning” to reveal the semantic meaning of concepts), which was required to be trained with some simplifying assumptions. First of all, the number on the image must be exactly present, and the image must be prepared in such a way that the required number occupies about one third of it. Also, the team quite reasonably admitted that the length of the house number can be limited to five digits, which is acceptable for most urban numbering systems in the world.
Here is the raw source for the neural network:
')
The network training took about six days and was carried out on the materials of the publicly accessible Street View House Numbers database, which already contains approximately 200,000 of the same numbers. As a result, after learning, the recognition accuracy of the Google system was 96%, which is comparable to the human indicator of similar work at 98% - this figure will be the goal for further research.
[ Source ]
The work is called Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks and it is about using neural networks for recognition purposes. As is obvious, the case is greatly complicated by many factors: the different locations of the house number, the color of the plate, its various inclinations, the quality of the plate itself, as well as its photographs, and a number of others.
As a result, the task was shifted to the modification of the DistBelief neural network with 11 levels of neurons (this is the same network that Google uses for “deep learning” to reveal the semantic meaning of concepts), which was required to be trained with some simplifying assumptions. First of all, the number on the image must be exactly present, and the image must be prepared in such a way that the required number occupies about one third of it. Also, the team quite reasonably admitted that the length of the house number can be limited to five digits, which is acceptable for most urban numbering systems in the world.
Here is the raw source for the neural network:
')
The network training took about six days and was carried out on the materials of the publicly accessible Street View House Numbers database, which already contains approximately 200,000 of the same numbers. As a result, after learning, the recognition accuracy of the Google system was 96%, which is comparable to the human indicator of similar work at 98% - this figure will be the goal for further research.
[ Source ]
Source: https://habr.com/ru/post/208330/
All Articles