Security is a major concern in airports, government buildings and major infrastructures around the world. Governments need to be able to quickly identify potential threats among the many people that enter and exit their countries daily. An effective facial recognition system is critical in safeguarding the country and critical infrastructures.
Panasonic has been at the forefront of facial recognition for a long time and its surveillance system and vision systems are used around the world.
The Japanese company has a strong research and development (R&D) presence in Singapore that focuses on using artificial intelligence (AI) and deep learning to develop such systems to improve security, safely and intelligence.
“While our focus is to develop and support internal Panasonic businesses, we are also opened to working with partners to provide our solution in Singapore and the region,” said Dr Pongsak Lasang, Team Leader, Imaging Processing & 3D Team, Core Technology Group, Panasonic R&D Center Singapore.
Founded in 1990, the centre is a wholly-owned R&D arm of Panasonic.
In Singapore, it is working with the government on the Smart Nation initiative, which aims to harness digital technologies to build a future Singapore, improve living and build a closer community, empower citizens to achieve their aspirations through good jobs and opportunities, and encourage businesses to innovate and grow.
Panasonic’s involvement is to provide deep learning technology in the areas of detection and recognition.
Machine learning roadblock
“We started working on developing a facial recognition solution using traditional machine learning 10 years ago. In 2012, we began to understand the limitations of machine learning when we had difficulty deploying systems in different situations. We needed to adjust our model and sometimes, our system was not robust enough,” recalled Lasang, who leads a team of more than 80 researchers focused on deep learning and machine learning.
The breakthrough came in 2014 with the advent of deep learning which allowed the R&D team to capture and utilise massive amounts of complex data needed for facial recognition.
Panasonic began putting together multiple NVIDIA Titan X GPUs to set up GPU servers to help in its R&D efforts. This allowed the team to improve the robustness and deployment of the system to become more adaptable to different situations.
Lasang is appreciative of NVIDIA for creating a computation platform and system with multiple GPUs, high performance computing and many ports that can accelerate tasks and training.
“We need to adjust the parameters and big data, then we need to train and analyse in a short time. Our two NVIDIA DGX systems allow us to achieve this target. When we utilised these systems, we could figure out and train different parameters quickly,” he said.
With the two NVIDIA deep learning supercomputers, Panasonic was able to develop the world’s first facial recognition solution.
It took part in the International Conference on Computer Vision (ICCV) Face Challenge and came up tops in two categories.
Deep learning success
Building on the success, Panasonic announced in February 2018 that it will be release face recognition server software using deep learning technology in July 2018 outside Japan, and in August 2018 in Japan.
The Face Recognition Server Software WV-ASF950 and the Face Registration Expansion Kit WV-ASFE951W will feature a core engine that boasts what is claimed to be the world’s highest face recognition performance. This high-precision face recognition software can identify faces that are difficult to recognise with conventional technologies, including faces at an angle of up to 45 degrees to the left or right or 30 degrees up or down, and those partially hidden by sunglasses.
Good NVIDIA experience
“We have had a good experience using NVIDIA GPUs and are interested in the DGX 2,” said Lasang.
Announced at GPU Technology Conference in San Jose in March and expected to be available later this year, the NVIDIA DGX 2 is the first 2 petaFLOPS system that combines 16 fully interconnected GPUs for 10 times the deep learning performance, enabling researchers to work on complex AI challenges.
“We have many projects on hand and our two DGX supercomputers and four GPU servers are simply not enough,” Lasang added.