The software Google says that in that picture out “a group of young people playing frisbee”
The program is now able to describe a complex scene of a photograph accurately
grammatically correct sentences
- Thursday, November 20, 2014
- By Tom Simonite
- In Francisco Reyes
Photos: An experimental software of Google is able to accurately describe scenes photos, like the two on the left. Although still make mistakes, as seen in the two photos on the right.
Google researchers have created a complete software able to use to describe accurately scenes shown in photos sentences, what which represents a significant advance in the field of computer vision. When shown a picture of people playing frisbee , for example, the software responded with the description “A group of youths playing frisbee “. The software can even count, giving answers like “Two pizzas placed in an oven”.
So far, most efforts to create software that understood the images were focused on identifying individual objects, a task that is easier.
“It’s very exciting,” says research scientist at Google, Oriol Vinyals. “I’m sure that will some potential applications of this project.”
The new software is the latest Google research in the use of large collections of simulated neurons for processing data (see “deep learning”). Nobody in the new Google software programmed with rules on how to interpret the scenes. Instead, their networks “learned” after consuming data. Although according Vinyals for now is only a research project, he and fellow Google have already started thinking about how to use it to improve image search or help the visually impaired to navigate online or in the real world.
Researchers at Google created the software through a type of digital brain surgery, connecting together two neural networks developed separately for different tasks. A network had been trained to process images and create a mathematical representation of its contents, thus preparing for the identification of objects. The other had been trained to generate complete sentences in English as part of automatic translation software.
When networks are combined, the former can “see” an image and then provide a mathematical description of what ” sees “the second, it uses that information to generate a readable sentence. The combined network was trained to generate more accurate descriptions. For this they were tens of thousands of images with descriptions written by human beings. “Through language we see what we thought was the image,” says Vinyals.
After this training process, the software was used with several large datasets of images from Flickr and other sources, and asked that describe. Then the accuracy of their descriptions are judged by an automated test used to determine the performance of computer vision software. Google software got scores in the range of 60 on a scale of 100 points. Humans who test normally achieve a score in the range of 70, according Vinyals.
This result suggests that Google is far ahead of other researchers dedicated to creating software for describing scenes. Recently, Stanford researchers have published details of their own system, ensuring that scored between 40 and 50 on the same standard test.
However, Vinyals notes that Google researchers and elsewhere are still in the early stages of understanding of how to create and test this type of software. When humans Google asked to rate their software descriptions of images on a scale of 1 to 4, the averaged was only 2.5, suggesting that still has a long way to go.
Vinyals predicts that research on understanding and description of scenes will intensify. One problem that could cause delays is that, although they have created large databases of images labeled by hand to train the software to recognize individual objects, fewer photos tagged more natural scenes.
year, Microsoft launched a database called COCO to try to fix this situation. Google used in its new research COCO, but still relatively small. “I hope other partners contribute to improve,” concludes Vinyals.
No comments:
Post a Comment