ETRI on-device vision processing software,
Korea's first acquisition of the Khronos Group OpenVX standard
Increased low-power performance by applying GPU parallel computing operations


On-device vision processing software developed domestically has received international standard certification.

The Electronics and Telecommunications Research Institute (ETRI) announced on the 26th that it has developed high-performance, low-power vision processing software for on-device devices and obtained OpenVX standard compliance certification from the Khronos Group, an international standards consortium.

With intelligence emerging as a hot topic across all industries, demand for real-time on-device monitoring of on-site situations is growing. However, simultaneous utilization of existing vision processing and AI technologies requires enhanced performance.

Furthermore, because each on-device device's SoC computing resources differed, companies had to develop software tailored to their individual products, resulting in severe fragmentation. Furthermore, software development required significant expertise, technical expertise, time, and cost, posing challenges in industrial settings.

Computer software tends to be well-compatible, no matter which company you buy it from. On the other hand, smartphones require software development and optimization each time because the hardware chips and characteristics used by each manufacturer and new product are different.

The OpenVX-based results developed by ETRI can operate on a variety of hardware with a single application software development. They also feature an automatically optimized execution environment. This is expected to be a significant contribution to fundamentally resolving software portability and compatibility issues in on-device devices.

ETRI passed a total of 6,162 functional tests and became the first in Korea to receive OpenVX certification. This technology is expected to elevate the level of machine vision in various industries, including smart mobility, smart factories, autonomous robots, and drones.

The research team selected a vision module that fits the product hardware environment according to the standard, connected it, and made it possible to automatically optimize performance. By creating modules using the best coding methods, we can improve performance without having to code each function individually.

The research team also developed a lightweight runtime environment for running OpenVX applications. This technology allows for the parallel computing capabilities of mobile GPUs while still complying with the OpenVX standard.

Depending on the usage environment, computing resources can be optimally utilized for calculations. The research team confirmed that using a GPU in combination with a CPU alone can increase performance and reduce power consumption.

In the future, the research team plans to expand this technology into a comprehensive on-device AI computing software platform that encompasses the entire vision recognition processing process, from data preprocessing to deep learning.

This research was conducted as one of the projects of the Ministry of Science and ICT's 'On-device Intelligent Information Processing Acceleration SW Platform Technology Development for Smart Devices.'