A computer program that can identify the apparent age of anonymous individuals from images.
Automatically recognize the actors in a movie or the sportspeople in a televised competition, establish the age of a viewer in front of a screen – for example for better parental control – or even identify customers via a selfie as tested by the HSBC bank… everyday practical visual recognition applications are on their way. But the research is taking great leaps forward and Orange, which has been working on it for a decade, is now at the forefront.
Last spring, a team of researchers based in the Cesson-Sévigné Orange Labs, near Rennes, won the ChaLearn challenge organized as part of the annual Conference on CVPR (Computer Vision and Pattern Recognition). The challenge? Using deep learning technologies, ask a computer program to identify the apparent age of anonymous individuals from images.
VP Data & Knowledge Research at Orange, Henri Sanson explains: “A dozen human testers first annotated the faces in a collection of photos, giving their perceptions of their ages. The algorithms then have to predict the apparent age closest to the average age given by the panel of testers.” However, the difficulties in this visual recognition are proportional to the age of the person in the photo. “Some people may be fifty and look forty!” adds Henri Sanson.
But even before precisely analyzing the characteristics, eyes, mouth, nose, to come up with an age, the computer program must first successfully… detect that there is a face in the image. “Finding faces in images may seem really simple because we all have it on our smartphones,” notes Henri Sanson, but in reality it is no small matter. The work of researchers and technological innovation mean that detecting a face on an image is now more reliable. “Technology has made great strides in the last two years,” says Orange Data & Knowledge Research VP. An acceleration linked to the progress made in artificial intelligence and in particular convolutional neural networks – which Orange was among the first to work on and which represent the gold standard in terms of visual recognition.