AI iris recognition technology is shifting from the cloud to end devices.
In the past, high-precision AI iris recognition often relied on cloud computing power, which created inherent application bottlenecks in scenarios with limited network access or strict privacy requirements. How to migrate algorithm capabilities to edge devices while maintaining recognition accuracy has become a common technical challenge for the industry.
Homsh has provided its own solution.
I. Domestic NPU Platform: Enabling End-to-End Iris Recognition
Recently, the R&D team of Homsh completed the deployment and verification of the company's self-developed lightweight iris recognition model on an embedded NPU platform. The target hardware selected is a development board based on the Rockchip RK3588 chip—a representative product of high-performance domestic edge computing chips, adopting the aarch64 architecture and equipped with a dedicated NPU computing unit.
The team systematically verified two technical routes on this platform: a general inference solution based on ONNX Runtime and an NPU acceleration solution based on RKNN. Both routes have completed model loading, inference link connection and function verification, and the supporting graphical user interface can normally perform offline evaluation and real-time camera capture.
This means that Homsh's core iris recognition algorithm now has the ability to run independently on domestic edge chips.
II. 3.7x Speed Improvement: Remarkable NPU Acceleration Effect
Performance data provides the most intuitive illustration.
Under standard test conditions, the ONNX model achieves 100% iris recognition accuracy with a stable inference frame rate of approximately 1 FPS. In contrast, the RKNN model accelerated by NPU sees its inference frame rate jump to 3.64 FPS, representing a speed improvement of about 3.7 times.
Behind this performance leap is the team's success in overcoming multiple technical obstacles, including RKNN model export, underlying library architecture compatibility, and missing symbol definitions. From algorithm transplantation to hardware adaptation, every step has verified the maturity of Homsh's vertical integration capabilities in "algorithm—chip—terminal".
Currently, the team is conducting further research on the accuracy optimization of the RKNN model, aiming to restore the recognition accuracy to a level comparable to the ONNX version while maintaining the advantage of high frame rate.
III. Edge Deployment: Unlocking More Application Possibilities
The value of edge intelligence goes beyond speed alone.
When iris recognition capabilities are integrated into a small development board, it breaks free from the reliance on cloud computing power and stable network connections. For scenarios with limited network conditions such as underground mines, remote construction sites, and mobile law enforcement, this means a truly implementable solution.
At the same time, in edge deployment mode, biometric data can be matched without being uploaded to the cloud, which is naturally suitable for application scenarios with strict data security requirements such as financial outlets and government services.
Homsh will continue to promote the in-depth adaptation of lightweight algorithms to domestic edge chips, providing partners with high-performance, low-cost, and easy-to-integrate edge solutions for iris recognition.
IV. Technical Highlights
Quick Overview of Technical Highlights
Target Platform: Rockchip RK3588
Model Type: Iris Recognition + Face Recognition
ONNX Accuracy: 100%
RKNN Frame Rate: 3.64 FPS
Function Verification: Offline evaluation, real-time capture, 1:N recognition mode
From algorithm R&D to chip adaptation, from cloud deployment to edge implementation, Homsh is expanding the application boundaries of iris recognition technology step by step.
Making recognition faster, closer, and safer.
For more technical details or to discuss cooperation opportunities, please feel free to contact us.
