The hardware and software for AI acceleration
If you develop facial recognition software, you spend most of your time and effort maximizing performance, accuracy and precision. Your customers expect high performance, even when running on low-power edge devices like point-of-sale (POS) terminals, kiosks and digital signage. But customers also expect high accuracy and precision, with false-match rates of less than one in 10,000 or better, and anti-spoofing capabilities against presentation attacks using printed photos, videos and silicone masks.
Given those stringent requirements, the AI models that CyberLink has built into their FaceMe product are more compute-intensive than many other edge AI algorithms. To run their models on edge devices with the high performance, accuracy and precision their customers demand, CyberLink has ported FaceMe to run on processors from Qualcomm Technologies, Inc.
These processors offer the computing power necessary to run facial recognition applications at low power consumptions, allowing for a wide range of device form factors. Using the Qualcomm Neural Processing SDK for AI, CyberLink optimized the performance of FaceMe on the Qualcomm Hexagon neural processing unit (NPU).
The hardware and software for AI acceleration
CyberLink FaceMe is ranked among the top vendors worldwide in the Face Recognition Technology Evaluation (FRTE) by the National Institute of Standards and Technology (NIST).
That includes ranking number 1 (excluding China/Russia vendors) in the NIST FRTE 1:N VISA-Border category, with 99.73% accuracy. And in the NIST FRTE 1:1 VISA category FaceMe ranks number 4, with a True Acceptance Rate of 99.83% and a False Match Rate of 1 in a million (1-6).
The software supports major operating systems, AI inference engines and edge computing platforms, and is optimized to leverage hardware acceleration available on processors. When CyberLink looked for ways to maximize performance, precision and accuracy, they turned to hardware acceleration on the Hexagon NPU and Qualcomm Adreno GPU available on Qualcomm Technologies’ system-on-chips (SoCs).
That led them to use the Qualcomm Neural Processing SDK for AI, supported on most Qualcomm Technologies’ SoCs as the main software tool for porting the FaceMe facial recognition AI models. The SDK enables developers to optimize and deploy their AI models on Qualcomm AI products.
Porting CyberLink FaceMe to SoCs from Qualcomm Technologies
CyberLink engineers followed the model workflow described in the documentation:
Figure 1: Qualcomm Neural Processing SDK workflow
- Model conversion and initial verification – Starting with a trained model file, the engineers used the tools in the SDK to convert their AI models to Deep Learning Container (DLC) format. Using SDK tools on Ubuntu, they verified the model outputs against expected behavior.
- Integration with Neural Processing SDK for model inferencing – They enabled the SDK in FaceMe by integrating the SDK as a new inference engine within their application. They then verified that the converted models ran correctly in CPU and GPU modes, with expected performance, precision and accuracy.
- Model quantization and optimization – The engineers quantized all their models for INT8, then ran them on NPU and verified accuracy and stability.
- Benchmarking and validation – They benchmarked and validated to measure inference speed and accuracy across different SoCs from Qualcomm Technologies. They ensured that all models were of shipping quality, balancing performance, precision and accuracy for real-world deployment.
CyberLink went through the process of model conversion, quantization and engine integration. Quantization turned out to be the more involved task, requiring fine-tuning and validation to achieve the high accuracy required for their application.
CyberLink found the process of converting models smooth and straightforward. Being able to test converted models on Ubuntu made the initial validation efficient and convenient.
It took more work to regain accuracy through model quantization and tuning. The SDK offers multiple quantization methods, each with varying trade-offs in accuracy and precision. With a high bar for accuracy, finding the optimal method required extensive testing to identify the process that resulted in the smallest drop in accuracy.
The SDK and tools enabled the engineers to achieve the desired optimization on their own, with Qualcomm Technologies’ experts providing guidance.
Testing and QA
CyberLink has tested its ported FaceMe application on several SoCs from Qualcomm Technologies, including the Qualcomm Dragonwing QCS6490|QCS5430 and Qualcomm SM6225|QCS6125|QCS610 and other Qualcomm Technologies platforms. The testing has demonstrated the success of the project by achieving the target accuracy and performance.
CyberLink provides an SDK so customers can integrate facial recognition into their software and hardware products, and the SDK includes sample code and demo systems. The company’s QA team has successfully tested sample code and demos with the ported application on some of CyberLink’s customers’ devices and on development kits powered by processors from Qualcomm Technologies. Testing has gone smoothly, with the QA team finding it easy to learn, use and validate the ported application.
Results
According to CyberLink’s measurements, the accuracy and precision of the FaceMe application remains consistent across different processors. The difference lies in performance.
The following table summarizes performance results for pure face recognition:
| Platform | Performance (frames per second) |
| Vendor A on CPU (4x Cortex-A53, 1.8 GHz) | 4.2 |
| Vendor A on NPU | 13.0 |
| Vendor B on APU | 16.0 |
| Qualcomm QCS6125 on Adreno GPU | 12.2 |
| Qualcomm QCS6125 on Hexagon NPU | 33.8 |
| Snapdragon 870 on Adreno GPU | 62.0 |
| Snapdragon 870 on Hexagon NPU | 87.0 |
Biometrics in IoT
To ensure user privacy and security, CyberLink complies with a range of region-specific privacy laws and biometric regulations including GDPR (Europe) and CCPA (California). It also adheres to other, nation-specific laws for protecting biometric data.
The company conforms to standards set by ISO, including the following:
- ISO/IEC 27001 (Information Security Management System) to ensure data security
- ISO/IEC 27701 (Privacy Information Management System) to ensure compliance with privacy regulations and the protection of personally identifiable information (PII)
- ISO/IEC 30107 (Presentation Attack Detection) to strengthen anti-spoofing measures in biometric systems
CyberLink educates prospects on the benefits of FaceMe, while addressing privacy and security concerns, in three ways. First, their AI models and software are proprietary – developed entirely in-house by their R&D team in Taiwan. Unlike solutions that rely on third-party datasets or outsourced AI models – potentially raising privacy concerns – the company ensures full control over data handling and compliance.
Next, the primary use cases for FaceMe, including authentication (services and apps), electronic Know Your Customer (eKYC), two-factor authentication (2FA), access control and time attendance, require explicit user consent during enrollment. That means that users actively register themselves and are fully aware of and agree to the use of facial recognition, in compliance with privacy regulations.
Finally, CyberLink emphasizes that biometric data must be handled with extra care. Their technology and customer implementations follow strict security practices, including data encryption, secure transmission protocols and compliance with industry standards. Their system is designed so that biometric data and personal information can be completely deleted upon request, ensuring that users have full control and privacy.
In the biometrics industry, customers demand robust, secure solutions. CyberLink believes that trust – including privacy compliance, reliability in vendor collaborations and an emphasis on security – is an important differentiator in Qualcomm Technologies’ products. Adding to core values, CyberLink appreciates that quality, stability, brand reputation and strong sales channels are also important to customers, making Qualcomm Technologies a logical choice for biometric applications.
Next steps
CyberLink’s customers run FaceMe on IoT devices ranging from high-end workstations and PCs to access control devices and kiosks on the edge (see below).
Come for support, stay for the community
Get support from experts, connect with like-minded developers, and access exclusive virtual events.
Figure 2: CyberLink’s FaceMe application/SDK – Architecture
By porting their application to SoCs from Qualcomm CyberLink has implemented AI acceleration on NPU and GPU, with excellent results in processing speed. The result is accurate, precise, real-time facial recognition with low latency, low power consumption and on-device. Learn more about applications for FaceMe in security, access control, eKYC, finance, 2FA, retail/hospitality and robotics/manufacturing.
See how you can use the Qualcomm Neural Processing SDK for AI to accelerate your AI workloads on processors from Qualcomm Technologies.
Browse the documentation, then visit our Developer Discord for deeper insights and real-time conversations with fellow developers and Qualcomm Technologies’ experts.
Qualcomm Blog. V. H.
Contacto