LSFM model technology can capture suspects from other angles

Scientific and technological developments, according to foreign media reports, I do not know whether you've used Snapchat phone software, it turns your photos into a bear, or blend your face to face and others. A group of researchers recently invented the most advanced technology for building 3D facial models on computers today. The system can improve personal avatars in video games, enhance the accuracy of secure face recognition, and enable Snapchat to introduce more filters.

LSFM model technology can capture suspects from other angles

When a computer processes a face image, a so-called "3D deformation model" (3DMM for short) may be used. This model represents the average level of the face, but it also contains common feature information that deviates from the average level. For example, if you have a long nose, your two jaws may be longer. Considering this type of correlation, when the computer categorizes your facial features, instead of storing every pixel of the 3D scan, you can list hundreds of different values ​​that describe your average facial shape, including The parameters related to age, gender and face length are generally relevant.

However, one thing needs attention. In order to record various facial differences, the 3D deformation model needs to integrate information of many faces. So far this has required scanning many faces and then listing all facial features. The current best performing model also contains only a few hundred people's facial information. Most of them are adult white people and therefore have limited levels when dealing with other age and ethnic groups.

James Booth, a computer scientist at Imperial College of Technology, and colleagues introduced a new technology that allows 3D models to be built automatically and covers a richer population. This method is mainly divided into three steps. First, the algorithm uses an algorithm to complete the automatic face scan, and labels the nose tip and other feature points. Second, another algorithm is used to sort all scan results by features and integrate them into a model. Finally, use the third algorithm to find and remove invalid scan results.

"The important contribution of this research work is that researchers demonstrated how to automate this process," pointed out William Smith, a computer vision expert at York University. He did not participate in this study. Alan Brunton of the Fraunhofer Institute for Computer Graphics Research in Darmstadt, Germany, also stated that the labeling of facial features of a large number of faces is a " Extremely tedious work. “You think that labeling is nothing but tapping facial features, but it's not that easy. Sometimes it's difficult to see where the mouth is, so even if it's purely artificial, it's hard to avoid errors.”

But Booth and colleagues did not stop there. They applied this method to a set of nearly 10,000 facial scans with significant demographic differences. The scans were collected by orthopaedic surgeons Allan Ponniah and David Dunaway at a science museum in London. They hoped to improve the outcome of reconstruction surgery. So they found Stefanos Zafeiriou, computer scientist at Imperial College, to help them analyze the data.

They applied the new algorithm to these scan results to create a "large-scale facial model" (abbreviated as LSFM). When tested simultaneously with the existing model, the accuracy of the LSFM model is much higher. In one of these comparisons, they used a child's photo to create the child's facial model. The model created using LSFM looks very close to the child, but the facial models generated by most of the existing deformed models appear to be only an unrelated adult because the latter is based on adult data. Booth and colleagues even collected enough scans to create a more accurate deformation model with people of different races and ages. Their model can automatically divide their age range based on facial shape.

Booth's team has put this new model into use. In another paper, researchers used 100,000 human faces synthesized by the LSFM model to train an artificial intelligence program to convert ordinary 2D snapshots into an accurate 3D model. This technology can help us to observe the suspects captured by the camera from other angles, or predict their looks after 20 years. We can even use historical portraits to create lifelike character models.

The LSFM model also has medical application value. If someone has a nose defect, the technique can help the plastic surgeon to predict the appearance of the new nose based on other facial features. Facial scans can also be used to diagnose patients with Williams syndrome, which is often associated with heart disease and developmental delay, as well as facial features such as a short nose and wide mouth. More accurate facial models and difference information can increase the sensitivity of such tests. As Pornia pointed out, this new model "opens a new door for us."

Next, the researchers are ready to add facial expressions to the model, whether you are frowning or smirking, can be identified. Zaffirier said that they will return to the museum to scan the faces of more tourists.