Over the past decade, as we have made great strides in artificial intelligence, we have been able to add advanced capabilities, such as face recognition, to embedded systems. Despite the many benefits of facial recognition, its use is still sometimes viewed as problematic and even controversial. How are the facts? In this article, we will clear up some misconceptions about face recognition.
1) The cost of face recognition is very expensive
One would feel that for computers to be able to recognize faces, the solution would have to be high-end hardware. After all, since the mid-2000s, breakthroughs in image classification by deep learning algorithms have harnessed the processing power of graphics processing units (GPUs), which often used in tightly coupled clusters. But for developers of facial recognition applications for embedded systems such as home security and access control products, such a complex machine learning process not required. Designing efficient algorithms that focus on detecting faces and matching them to registered images will require far less processing power than research-level computing power.
2) Face recognition is very difficult
A key challenge in machine learning is matching the design process to the application so that it can produce useful results when training. But in applications such as face learning, there is no need to build these structures from scratch. We can use platforms built on proven machine learning processes that not only deliver high performance quickly but also offer a degree of customization to meet the needs of different target markets.
3) Face recognition requires high-performance processing
Many people see that in a cloud computing environment, we use high-performance hardware for machine learning, so they take it for granted that machine learning is a heavyweight process. However, these systems need to be adaptable to many different applications, and they can take advantage of open source tools that support all deep learning architectures.
Thus, even for inference applications, when using the network to analyze real data, the model has a high degree of data and computational redundancy. Embedded solutions can significantly reduce these overheads, enabling complex face recognition algorithms to run on 32-bit MCUs.
4) Face recognition is not very secure
An important application of face recognition in embedded systems is access control. There are concerns that some criminals are trying to use masks; photos or videos in an attempt to trick facial recognition devices. In fact, the industry has long since upgraded face recognition, incorporating a live detection algorithm; which can determine whether the submitted biometrics come from living individuals based on information such as head movement; breathing, and red-eye effects. Secondly, there is cooperative living detection, which requires the user to make corresponding actions according to the prompts; and verify whether the user a real living body through cooperative combined actions such as blinking; opening the mouth, shaking his head, and nodding. Liveness detection technology effectively prevents malicious persons from forging and stealing other people’s biometrics for identity authentication.
5) Face recognition violates privacy
Many applications familiar to the public require uploading raw data to a cloud server, where the data then processed. This is a concern for many consumers, who do not want their activities in and around their homes to spread over the Internet; possibly even after a malicious attack on their servers. Good news is that there are some platforms; such as the Macrosafe face recognition solution, perform all image processing and face recognition locally; and data never leaves the platform, ensuring that the final product protects the user’s privacy to the greatest extent possible.
6) Face recognition cannot be performed in the dark
There are situations where speed gates with integrated facial recognition are required to operate in less than ideal lighting conditions. Facial recognition technology seems to rely on visible light to work properly; and night work or power outages can be a problem. However, this problem can solved very simply by using a visible light image sensor with an auxiliary device that operates on the infrared spectrum, or using time-of-flight data to build a 3D map of objects in range. With this approach, the absence of light is no longer a problem; and it also helps improve usability and reduce power consumption by not requiring the solution to employ artificial lighting.
7) Face recognition requires expertise in artificial intelligence
Overall, artificial intelligence is a very broad and complex field. In deep learning alone, new academic papers appear on the internet every day; exploring different technical fields and new pipeline structures. Since the development of recognition solutions and equipment; especially after the Covid epidemic has catalyzed a large number of demands; major companies have increased their investment and development in face recognition technology.
Today’s face recognition equipment is very mature and sound. The Macrosafe recognition terminal can used to detect masks and safety helmets; and can also integrate a temperature measurement module to perform dual detection of recognition; and body temperature detection on the passers-by of speed gates at the same time.
8) Face recognition equipment can only be used indoors
When the face recognition terminal was first launched on the market; it was indeed only used indoors or outdoors with a canopy. However, when used in outdoor and open-air situations, the face recognition terminal needs to have a good waterproof function to avoid problems such as corrosion and short circuit of the circuit board due to weather during use. Macrosafe speed gate with face recognition terminal has passed the IP67 waterproof test, and can be used outdoors without a canopy.
Speed gates with recognition camera are not only an ideal solution for indoor access control but also the perfect choice for the outdoor entrance control of school gates, construction sites, and scenic spots
9) Training is a tedious task for the end user
Early face recognition implementations in embedded systems such as tablets and smartphones required a range of different poses in order to efficiently train neural networks for recognizing the faces of new users. With advances in techniques such as transfer learning; it only takes one face-to-face camera to train features and add them to a database of permission users.
10) The application of face recognition is limited
Like any technology, it’s hard to imagine how the technology will be used until innovative companies put it to work. Face recognition combined with security speed gates for building entrances or checkpoints is a common way to use it. It can reduce labor costs while verifying the identity information of interviewers more efficiently and intelligently, preventing unnecessary access.
Beyond that, smart appliances and power tools can use this technology for safety purposes; disabling features to prevent injury to children. Devices will no longer be designed to recognize faces, but also expressions. Devices can read emotional signals, such as disappointment, confusion or joy, and respond accordingly, improving the overall user experience.