“With the help of deep neural networks, people researchers at the Russian State Higher Economic University have proposed a new method that can identify people’s identities from videos. This method does not require a large number of photos, and has significantly higher recognition accuracy than existing methods-even if only one photo of a certain person is available.
With the help of deep neural networks, people researchers at the Russian State Higher Economic University have proposed a new method that can identify people’s identities from videos. This method does not require a large number of photos, and has significantly higher recognition accuracy than existing methods-even if only one photo of a certain person is available.
Facial recognition technology has developed rapidly in the past few years. As tools to verify and identify individuals, these technologies are used in a variety of fields, from law enforcement agencies that fight terrorism to social networks and mobile applications. Many international companies and research teams from leading universities in the world are constantly experimenting with data and the instrument itself to improve the accuracy of recognition.
Recognition can be done in many ways, but the best results have recently been obtained with the help of high-precision neural networks. The more training images the neural network obtains, the more effective the process will be. The network can extract key facial features and then use this knowledge when identifying unknown images.
Now, it is easier to access more and more photo data sets and use these data sets to train neural networks. For restricted observation environments (photos with the same face orientation, lighting, etc.), the accuracy of the algorithm has already reached the level of human facial recognition capabilities. However, under the condition of unconstrained recognition, the collected video data has variable illumination, angle and size, and achieving high-precision recognition is a greater challenge for researchers.
Savchenko, a professor in the Information System and Technology Department of the National University of Economics, explained: “The network can identify well-known actors with 100% accuracy because the number of available images of actors may be millions. However, as the knowledge accumulated in the neural network This does not mean that it can adapt to the situation where only one photo is used as a training sample and identify a person’s identity.”
To solve this problem, researchers from the National University of Economics use fuzzy sets and probability theory to develop video recognition algorithms. In experiments using a small number of images for video real-time facial recognition, the algorithm significantly improved the accuracy of several well-known neural network architectures (2-6% higher than earlier experiments), such as VGGFace, VGGFace2, ResFace and LightCNN.
As a test database, researchers from the National University of Economics used the following traditional data sets to evaluate video facial recognition methods: IJB-A (IARPA Janus Benchmark A) and YTF (YouTube Faces). These data sets contain free images of famous people (actors, politicians, public figures), and these images are collected from open sources in an unconstrained environment and at different points in time.
In the most complex experiment, the researchers used the above algorithm and several photos of the same person from another LFW (Labeled Faces in the Wild) dataset with better resolution to identify people from YouTube videos. The photos themselves were taken at different times (from the 1970s to the 2010s) and at different locations.
The essence of this new method is to use the relevant information of the reference photos, that is, the distance or gap between them. The connection between similar individuals (that is, the distance in the mathematical model) is smaller, while the connection between dissimilar individuals is larger. Understanding the degree of difference between people can help the system correct errors in the process of identifying video frames.
Professor Savchenko explained: “The algorithm estimates how close one frame is to one person and how close another frame is to another person. Then, it compares the similarity between the training still photos of the two people. Next, it Add the third person and evaluate which person he is more similar to-the first person or the second person. Then it will correct the recognition error.”
The algorithm has been implemented in Python language for fixed computers, which can help users find and group different people’s faces in photo/video albums, and estimate a person’s year of birth, gender, and other parameters. The researchers also developed a prototype Android application to determine the age and gender of people in photos and videos.
The analysis of the photo library can realize the automatic evaluation of the user’s social activity level and identify the user’s close friends and relatives. For modern smartphones, the prototype application is capable of processing 15 frames per second. According to the researchers, with their method, facial recognition can achieve higher accuracy.
The Links: LQ104S1DG2C CM800DU-12H LCD-STOCK