Now when people talk about face recognition, they often refer to visible light face recognition. Here I will no longer explore the traditional infrared face recognition solution, which is already a thing of the past.
Let’s first look at the types of face recognition access control solutions. There are three common solutions: pure CPU type, GPU type, and NPU type. Pure CPU type can be divided into several types: X86 type, ARM type, and single-chip type. The GPU type is usually a combination of weak CPU+GPU, such as X86CPU+GPU graphics card, or ARM+GPU SOC. The NPU type is usually a combination of weak CPU+NPU, such as ARM+NPU or MIPS+NPU SOC.
Each type has different performance and indicators because of the technical architecture. The advantages and disadvantages of each type of processor are shown in the figure below.
The two aspects of flexibility and isomorphism in the above figure are not very important in the field of face recognition, so the pros and cons of the face recognition solution mainly depend on the following three aspects, namely: performance (computing power), cost (overall cost ), power consumption (power consumption and heat generation).
Computing power represents the possibility of accuracy. The core indicators of face recognition are the library capacity and accuracy indicators. The library capacity is determined by the number of feature points in the model. The more feature points, the greater the computing power required.
The focus of cost is the comprehensive cost of the overall solution, not just the cost of the motherboard.
Power consumption is also a key indicator. In face recognition terminals, the power consumption is too high, and the application scenarios are often limited, and some scenarios are even unusable, such as battery-powered products.
Comparing CPU, GPU and NPU horizontally, NPU has advantages in three aspects. It can be predicted that SOC with NPU-based asic architecture will inevitably become the mainstream. In the past, CPU architecture, solutions such as: Intel's X86 solution, RK3288, RK3399 and other ARM solutions are all Gradually fade out of people's field of vision, only some low-cost single-chip solutions are still active in some ultra-low-end applications. However, due to the inherent high cost of the GPU architecture, it is only used in a small number of special scenarios.
NPU architecture, there are many solutions, the mainstream in the access control market are: HiSilicon 3516DV300, Novatek NT98525, Rockchip RV1109, sigmastar SSC338G, etc. In the door lock market, mainstream solutions include HiSilicon 3559V200, Novatek NT98566, sigmastar SSC337Q, Ingenic T31 (MIPS+NPU), etc. These two markets have different focuses. The access control market focuses on performance. Because it is access control, the requirements for the bottom library are relatively large. At the same time, false recognition and rejection of recognition must reach a certain accuracy. Because the door lock market is powered by batteries, the size of the base is not very big. The current solution only focuses on power consumption, and the performance is generally not high. Although the current schemes on the market focus on this design, it has also attracted controversy. Although the door lock has relatively small base requirements, it has high requirements for misidentification and refusal of identification. After all, this is the last hurdle for entering the house.