Visual Understanding and Cognitive Technology Gain Breakthrough Machine is Expected to “Read” Human Facial Expressions

The science and technology trends learned from Intel China Research Institute today that the hospital’s “leading edge visual understanding research” has achieved a breakthrough and the machine is expected to “read” human facial expressions.

It is understood that the research and development of this technology started in 2011. The first phase mainly focused on face detection and face recognition research; when researchers were able to distinguish between people's gender and age group, they began to enter the second phase, that is, Dense face key points, detection and tracking, and research on how to automatically recognize dynamic facial expressions; and in the third phase, the technology is further extended to 3D face analysis and deep learning.

Among them, based on the intensive facial keypoint detection and tracking technology, the researchers developed a set of facial beautification technology in 2015. In addition to dealing with the skin, the technology also incorporates some geometric processing - the change of face shape of the person. In the form that the user wants, for example, he can make his eyes bigger and his nose more firm.

Human emotion is a very important biometric information of interaction. Researchers said that in the field of human emotion recognition, facial expressions are mainly exercised by facial muscles, and muscle movement is related to a single expression. Different expressions are also relevant. At that time, the researchers mainly analyzed the relationship between different units by examining the muscle movement unit associated with a single expression. In 2015, the team won the champion of an American Computer Society competition. The competition was given. A lot of the pieces in the movie are marked with a main character to express his emotions.

This year, Intel’s team again called for this event based on deep learning algorithms. “A deep network, similar to a long tunnel, is basically invisible from the entrance. It's dark. So how do you make it transparent? Our plan is similar to playing in the middle and corner of the tunnel. A few holes, put a few searchlights, you can clearly see how the parameters of the network is passed.” Intel China Research Institute researchers said that through the idea of ​​polymerization supervision, single model recognition ability is relatively improved 5.5 percentage points. At present, this technology has exceeded the pass line in the classification accuracy of emotion recognition in a completely natural movie scene.

According to reports, researchers also trained multiple face recognition models from hundreds of thousands of face image data for intelligent analysis and 3D modeling of video images, which can not only detect and recognize faces but also accurately reconstruct 3D images. Face, and tracking changes in facial expressions in real time, attaching pre-designed special effects to 3D faces and superimposing them on the video to achieve realistic facial effects.