[ad_1]
Facebook researchers have presented a summary of a system they call "Rosetta", an automatic learning approach that boosts traditional optical character recognition, or "OCR", to exploit the hundreds of millions of photos uploaded to Facebook.
Suppose you want to look for memes in images on Facebook: the challenge of the site is to detect if there are printed letters in an image, then analyze these letters to know what a sentence says.
Also: Facebook accidentally exposes 14 million private messages from users
Of course, this technology has long been used for document processing, but the challenge for Facebook was to recognize text in any number of complex images, including text on an image or text on the Internet. . sign that was part of the original image, and then operate it at the scale of the constant flow of images from the site.
Facebook researchers Fedor Borisyuk, Albert Gordo and Viswanath Sivakumar shared the work on Rosetta at the Knowledge Discovery and Data Mining conference in London in late August, in an official article and two of the authors, Gordo and Sivakumar Manohar Paluri, proposed a A slightly simpler blog article describing the work.
Facebook has divided the task of "extracting" text from an image into two distinct points, namely to first detect if there is text in an image, then analyze what could be this word phrase.
Also: If Facebook worked, we would not be in this mess
For the first task, the detection, the authors used a convolutional neural network (CNN) called "Faster R-CNN", itself derived from the work done originally by Ross Girshick of Facebook when he was at Microsoft. Although CNNs have been used quite often over the last decade for image recognition tasks, such as ImageNet, R-CNN adds the notion of "regions" as a means of quickly identifying objects the object is located
Facebook has already widely deployed an object recognition system across its entire infrastructure called "Detectron", which clearly helped in this case.
Once the text is located in one image, the coordinates of that image are transmitted to another CNN to discern the word or phrase, character by character. The product of this second stage consists of character sequences constituting words and sentences.
Also: Facebook's new Rosetta AI system helps detect CNET hate speech
Because recognizing long words or long sentences can be particularly difficult, the authors describe the use of what is called a "curriculum" approach to form the character recognition system. They started by forming the system on small words of five characters or less and gradually increased the length of the words with the following iterations of the workout.
All the training work for the detection part and the recognition part were carried out using the "Caffe2" frame.
The authors devote a lot of time in the original article to describe how they tuned the system to optimize inference speed, when a new photo is examined and should be quickly searched for the text and the transcription. "Given our size and flow requirements, we spent [a] a significant amount of time to improve the speed of execution of the text detection model while maintaining high detection accuracy, "they write.
Also: How to protect yourself on Facebook TechRepublic
The Rosetta system is currently operational on the Facebook network, being used daily, write the authors. The authors propose that future challenges will include text extraction of video applications.
Previous and related coverage:
Should Mark Zuckerberg be fired? Security professionals have a say
Does the Cambridge Analytics scandal – and all the other privacy snafus on Facebook – justify the withdrawal of its CEO? A survey of security professionals gives them their opinion.
Facebook reveals new secret efforts to influence mid-term elections in 2018
The social network does not say who is behind the campaign, noting that it has no technical evidence at the moment to point fingers with confidence.
Facebook's "war room" hunt and destroy electoral interference, false news
The physical room will be responsible for protecting the network from insidious attempts to falsify mid-term elections in the United States.
Source link
