"Customer is God," and this proverb West reveals the objective laws of the customer occupies a central position of commercial activity. In order to better serve customers, optimize their business services and products, customer demand for analysis and research has been a top priority in business management analysis.
In the commercial Internet-based, digital society today, this law is more obvious. When ××× Web1.0 from the beginning, Cookie was invented for on the "customer" and filing digitized described normalized number and behavior of their Internet.
Web2.0 era followed, with the development of mobile Internet, personal life screen - computer interaction port changed much, the original way of "customer" digitized with Cookie has been unable to complete the full personal channels across all platforms screen behavior of normalization, in order to solve this problem, the device ID, SuperID came into being.
Currently, Web3.0 is in full swing of development, channels, platforms, terminals, the screen with the development of the IOT, by enabling cloud, sinking atomization of the "customer" to digitize the manner described ushered in a more complex and more Serious challenges. In view of this, the singularity cloud interested in this and many other vendors, many years ago began preliminary research for new forms of "customer ID" of. At present the basic consensus reached by two points:
IOT era, gradually digitized on-line physics community, for this open digital environment has been difficult sole limited by digital media equipment for customer behavior and digitized under the original line. This requires from individuals who will be directly digitized and extraction.
This new form of digital ID needs to be able to more efficiently and accurately pull through both digital information and original Web1.0 era of Web2.0.
After a period of trial and error, to face biometric-based Face ID scheme gradually reveal advantage, Face ID also become a major aspect of current singularities cloud business intelligence solutions for customers in the digital description. Based on this, StartDT AI Lab on face digitized direction to do a full and in-depth technical precipitation. Here's a few to show you:
Core Face recognition is naturally digitization, which contains digitized and accurate biometric people face comparison. As a visual intelligence engine integral part, StartDT AI Lab's face recognition technology can solve the problem of face recognition in complex scenes.
For example, under the dynamic video surveillance scene recognition, compared to the constraints people face at the scene face recognition technology to verify the required, one major challenge is the identification of unconstrained human face, face recognition of the difficulty lies in people Figure there will generally face blurred, obscured, great resolution, illumination and facial expression changes, these factors will affect face recognition to some extent, even significantly reduce the accuracy of face recognition. The StartDT AI Lab of recognition accuracy rate in this scene special issue of tackling major research and development using some of the following techniques:
01 data enhancement
When faced with the training data sample size is too small, uneven mass distribution or the distribution of the training set and the actual scene quite different, the generalization ability of the model will be a serious decline, then it makes sense to enhance the data; StartDT AI Lab at the same time by GAN network binding conventional image processing technology, the sample enhanced synthesis.
02 image processing
In the unconstrained scenes, is generally provided poor image quality, such as the human face, the resolution is generally poor, fuzzy, occlusion, and other low light, StartDT AI Lab depth in conjunction with conventional methods and learning methods, the human face image denoising, deblurring, such as super-resolution processing, resulting in higher quality images of the human face, to enhance the accuracy of the model the actual scene.
03 large-scale distributed parallel training
Training using multi-mode multi-machine cards, StartDT AI Lab currently supports millions of ID, hundreds of millions of photos the size of the training data set.
There is the so-called spear shield, there are anti-attack there. Ever since digital ID, there is a corresponding *** theft technology to break digital ID, refer the matter to the Face ID era still exist, and because of where Face ID is an open digital scene, *** means is rich easy again.
For example, using only photos on a cell phone or using a face transplant on APP theft of someone else's face in order to be certified, so it is very easy to be used by criminals, the scope of application of face recognition is significantly impaired. Therefore, we need to increase the in vivo detection face recognition were to be addressed before. Currently, the main method of face recognition *** including photos and video playback and three-dimensional mask *** ***.
We developed a variety of products in vivo detection methods for different application scenarios for unmanned retail scene, with the need for cross-validation of user-unfriendly way, and the need to control costs, so we developed a silent living RGB-based monocular Detection method. The main features extracted by the depth of learning styles and methods based on multi-feature fusion to the current 99.98% rejection rate scenario, 99.8% pass rate. At present the algorithm has been used in a variety of scenarios under our, guardians of our time face recognition system.
(Integrated in the product in vivo detection demonstration)
After the completion of the extraction Face ID as a natural extension of digital demand, video intelligence engine information on the human face also simultaneously digitized, such as age, gender expression and so on.
At present, people aged face major difficulties in forecasting is how to coordinate the continuity of the age, the age of the order among age segments of fuzziness, as well as from the actual scene makeup, the effect of light, angle and so on.
In the face of gender prediction, the main problem is that the intra-class variability, thus increasing Enhanced Data intra-class aspects of light, angle and other help to improve model performance.
In the facial expression recognition, difficulties encountered are mainly three aspects:
Each mode (lighting, posture, etc.) lack expression data set;
due to different factors such as age, gender, race, facial expression, etc., causing the intensity of high inter-subject variations;
Due to illumination, pose, occlusion caused by such factors large intra-class variability.
Age and gender expression recognition algorithm to predict the current StartDT AI Lab selected a great breakthrough in terms of solving the above problems, and through the training of large data samples, and achieved compared to the current market mainstream gender expression facial age more API high performance.
Using the above techniques show, I believe readers singular point cloud visual intelligence engine faces in the relevant technical capabilities have a certain understanding, "customer" in this digital era mainly described with Web3.0-based also on the Face ID You must know. From the current singularity cloud in practice Web3.0 perspective, Face ID at the head of 20% of high net worth clients VIP service to have full digital protection, which directly improve the business fell 80% expected return in the business model bags of ability. However, for the remaining 20% of the expected benefits, due to its decentralized business practices in the sparse audience of 80% in the long tail, how low-cost way to improve this part of the expected return of the pocket has always been the difficulty business scene. In view of this, the singular point cloud from a technical point of deconstruction of this part of the business scene, and time and time again through technological breakthroughs, continue to enhance the upper limit of the expected return. The technical details and the story behind this is the next issue of this column to share the theme, so stay tuned!