Break through face recognition application problems? Look at how security companies do [full text]

Scientific and technological developments in the past few years, there have been a number of security companies and the depth of the layout of AI released a variety of products, BrainChip be regarded as a security in the area of "new man", its main business is the development of AI and machine learning software and hardware, The pulsed neural network technology (SNN) has been used to achieve commercialization.

They introduced a new type of AI-driven software called "BrainChip Studio," which the company believes will change the way law enforcement and intelligence agencies use video surveillance to detect and classify people's faces. The software not only allows most traditional face recognition The solution can search a large number of video recordings faster and more efficiently, and it is not limited by some scenarios.

For example, BrainChip Studio can use low-resolution lenses without the need for high-definition video cameras for positive recognition; in addition the software can also be used for recording under poor conditions, it only requires a 24x24 pixel image to detect and Classified face.

Peter van der Made, founder of BrainChip, said: "Currently, convolutional neural network technology requires a large number of pre-marked data sets and expensive cloud platforms, and to achieve the same functionality, BrainChip's pulsed neural network technology only needs to be on the traditional CPU and internal Can be implemented in software. "

The number of global surveillance cameras is increasing. According to Lei Fengnet, it is estimated that 127 million surveillance cameras will be shipped this year, followed by the production of massive amounts of video data. This will require AI's effective analysis and rational use.

In a recent experiment, BrainChip's software was able to detect, extract, and classify three-and-a-half hours of video recorded by eight different cameras in real time, including over 500,000 face images; in another experiment, It can process 36 hours of video in less than two hours and extract more than 150,000 face images.

"We detected human faces from live or recorded video, and then created a pulsed neural network model for each face found," explained Robert Beachler, senior vice president of marketing and business development at brainchip, assuming it was taking a shot. In the game, we all know that this is the same person within the camera's entire field of view. After that, we can track the person’s dynamics (even if they change their position, clothes) within the camera's entire field of vision; in addition, the camera can also use these images. Divided into different modules, for any given camera (assuming 20), we can photograph a person walking through it, so that we have 20 images of our faces that we have photographed. “

"We have kept the cache of these facial pictures so that we can track them," he continued. "Once we have collected this information, we can compile these images into a database with a pulse model. If we find the suspect's image, run it through a pulsed neural network, create a model and compare it with the previously obtained image. Once it is found Suspect suspects, we will look at the person's image, and all information about the suspect will be placed.

How Pulsed Neural Networks Work

Unlike most AI solutions in the industry that use "convolutional neural network" technology to process data, BrainChip Studio uses "pulsed neural network" technology. According to Beachler, this is a computational approach that is more similar to the human brain and it works by mimicking the function of neurons and synapses.

"The neuron is made up of synapses connected to other neurons. The neuron itself is an integrated function of all the different inputs and then decides whether to trigger and pass the information to the downstream neurons. The transfer of this information is called a pulse, it actually Above is a bioelectric process. "This is a signal of a certain intensity through the synapses," Beachler explained. "The way the network is trained is through a threshold function that enhances or inhibits synaptic connections and changes neuron firing.

"The brain and computer work differently. Computers do mathematical problems. It's the same whether they are integers or decimals, but the brain doesn't think so," he added. “Humans are very good at recognizing patterns and knowing exactly what is, but humans are not very good at mathematics because our brain uses these pulse sequences as threshold logic to determine program programming.”

Therefore, when you understand these characteristics of human thinking in the form of a pulsed neural network and apply it to software, the computer will better recognize patterns in video images.

"Through this function, we know some interesting features of this type of pulsed neural network," explains Beachler. "One is that it is very good at using in a noisy environment; in addition, it only requires a single image, so unlike a convolutional neural network that requires a pre-marked set of information."

The advantages of pulsed neural networks

As mentioned above, compared to convolutional neural networks trained on data sets, pulsed neural networks require only a single image to learn and recognize patterns from them. "We can train on the spot to find patterns," added Beachler.

In addition, because pulsed neural networks do not require convolutional neural network mathematical operations, they require less computing power. Beachler quickly pointed out that the company is not just trying to create another type of biometric facial recognition solution, but can be used by video surveillance end users in everyday situations.

He said: "What we want to do is provide a product that can be used in the existing video surveillance environment. "We train our network in a deployment environment, we extract all faces in real time from the camera or recorded video. Information, we use the existing infrastructure, but it is not added to the camera or the attachment in the new DVR. “

The technology is currently used in casinos and police stations

Lei Feng Network (public number: Lei Feng network) learned that the company's current program has been applied in a large casino in Las Vegas. BrainChip partnered with SN Technologies to develop BrainChip-based SNAP technology for viewing the dynamics in the casino.

In addition, the company also cooperates with the French National Police Agency to reduce the time and manpower required by law enforcement officers to solve crimes.

Jean-Francois Lespes, head of Crimes Department of the Toulouse National Police, commented: “With this technology, law enforcement officers can quickly scan several TB of video footage, searching for features such as face, body shape and clothing patterns. BrainChip is based on pulsed neural networks. Technology can identify many types of objects and modes. This technology is very suitable for a noisy, low-resolution video surveillance system."