Shulou(Shulou.com)06/03 Report--

The western proverb "the customer is God" reveals the objective law that the customer occupies the central position of business activities. In order to better serve customers and optimize their own services and products, customer analysis and demand research has always been the top priority in business analysis.

In today's era of commercial Internet and social digitalization, this law is becoming more and more obvious. Since the time of Web1.0, Cookie has been invented to digitally describe and document "customers" and normalize their digital behavior under the Internet.

In the subsequent Web2.0 era, with the development of the mobile Internet, there are more screen-human-computer interaction ports in personal life. The original way of digitizing "customers" with Cookie has been unable to normalize the behavior of individuals across all channels and platforms. In order to solve this problem, devices such as ID and SuperID came into being.

At present, Web3.0 is in full swing, channels, platforms, terminals, screens with the development of IOT, through cloud empowerment, sinking fog, the way of digital description of "customer" has ushered in a more complex and severe challenge. In view of this, Singularity Cloud and many aspiring competitors began pre-research on the new form of "customer ID" many years ago. At present, there are two basic points of consensus:

With the advent of the IOT era, the original offline physical society is gradually digitized online. For this open digital environment, it has been difficult to digitize customers and behavior through the only limited media digital equipment. This requires direct digitization and extraction from natural persons.

This new form of digital ID needs to be able to efficiently and accurately pull through the existing digital information of the previous Web1.0 and Web2.0 era.

After a period of exploration, the Face ID scheme based on facial biometrics gradually shows its advantages, and Face ID has become the main technical scheme for the digital description of customers in the current singularity cloud business intelligence scheme. Based on this, StartDT AI Lab has made a full and in-depth technical precipitation in the direction of face digitization. Here are a few points to show you:

Naturally, the core of face digitization is face recognition, which includes the digitization and accurate comparison of facial biometrics. As an indispensable part of the visual intelligence engine, StartDT AI Lab's face recognition technology can solve the problem of face recognition in complex scenes.

For example, face recognition in dynamic video surveillance scenes, compared with the face recognition technology needed for face verification in constrained scenes, one of the major challenges is the recognition of unconstrained faces. The difficulty lies in the fact that the face images of face recognition generally have blurring, occlusion, low resolution, great changes in facial illumination and expression, and so on. These factors will affect face recognition to a certain extent. It even greatly reduces the accuracy of face recognition. On the other hand, StartDT AI Lab has carried out a special topic on the accuracy of face recognition in this scenario, mainly using the following technical means:

01 data enhancement

When the training data is faced with too small sample size, uneven quality distribution or great difference between the training set and the actual scene distribution, the generalization ability of the model will be seriously decreased, and the data enhancement is very meaningful; StartDT AI Lab uses GAN network combined with traditional image processing technology to enhance sample synthesis.

02 image processing

In unconstrained scenes, the image quality is generally poor, such as poor resolution, blur, occlusion, low light and so on. StartDT AI Lab combines traditional methods and depth learning methods to Denoise, deblur and super-resolution the face image, so as to obtain higher quality face images and improve the accuracy of the actual scene model.

03 large-scale distributed parallel training

Using the multi-machine and multi-card training method, StartDT AI Lab currently supports millions of ID and hundreds of millions of photo-sized training data sets.

As the saying goes, where there is a spear, there is a shield; where there is attack, there is defense. Since the advent of digital ID, there has been a corresponding * * technology to crack the embezzlement of digital ID. This problem still exists in the era of Face ID, and because Face ID is in an open digital scene, the means of * have become more abundant and convenient.

For example, using only a photo on a mobile phone or using a face-changing APP to steal other people's faces to be authenticated, it is very easy for criminals to use, and the scope of application of face recognition is greatly reduced. Therefore, we need to add live detection to deal with it before face recognition. At present, the main methods of face recognition include photo and video playback and stereo mask.

We have developed a variety of live detection methods for different application scenarios in our products. For unmanned retail scenarios, interactive verification is not user-friendly and needs to control costs, so we have developed a silent living detection method based on monocular RGB. Mainly through deep learning feature extraction and based on multi-feature fusion method to achieve 99.98% rejection rate and 99.8% pass rate in the current scene. At present, the algorithm has been used in a variety of scenarios to protect our face recognition system at all times.

(in vivo testing demonstration integrated on the product)

After completing the extraction of Face ID, as a natural extension of digital needs, video intelligence engine also digitizes face-related information simultaneously, such as age, gender expression and so on.

At present, the main difficulty of face age prediction lies in how to coordinate the continuity of age, the order of age, the fuzziness of age segmentation, as well as the influence of makeup, lighting, angle and so on.

In the aspect of face gender prediction, the main problem is intra-class variability, so increasing illumination, angle and other enhanced intra-class data will help to improve the performance of the model.

In facial expression recognition, there are three main difficulties:

Insufficient expression data set in various modes (illumination, posture, etc.)

High inter-subject variations is caused by age, sex, race, expression intensity and other factors.

Large intra-class variability caused by lighting, posture, occlusion and other factors.

At present, the age and gender expression prediction and recognition algorithm selected by StartDT AI Lab has made a great breakthrough in solving the above problems, and then through the training of big data samples, we have achieved a higher performance index than the current mainstream facial age and gender expression API on the market.

Through the above technical demonstration, I believe that readers have a certain understanding of the face-related technical capabilities of the singularity cloud visual intelligence engine, and also have a certain understanding of the main way of digital description of "customers" in the Face ID-based Web3.0 era. Judging from the current practice of Singularity Cloud in Web3.0, Face ID can be fully digitally guaranteed in the service of the top 20% high net worth VIP customers, which directly improves the merchants' ability to pocket 80% of the expected revenue in the business model. However, for the remaining 20% of the expected revenue, because it is scattered in the sparse business activities of 80% of the long-tailed customer base, how to improve this part of the expected revenue in a low-cost way has always been a difficult point in the business scene. In view of this, singularity cloud deconstructs this part of the business scene from a technical point of view, and constantly raises the upper limit of expected revenue through technological breakthroughs. The technical details and stories behind this are the theme of the next issue of this column. Please look forward to it.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.