Jinpin KG4204-G5 Helps Face++ Improve Face Recognition Speed

 

Background

A Beijing-based company called Face++ is an industry IoT solutions provider with artificial intelligence technology at its core. It is committed to providing leading AI algorithms and solutions for global industry users and building city-level intelligent IoT systems. Face++'s facial recognition technology, image recognition technology, intelligent video cloud products, intelligent sensor products, and intelligent robot products have been widely used in finance, mobile, security, logistics, retail, and other fields. Its core customers include not only industry-level leading enterprises such as Alibaba, Ant Financial, Huawei, and Lenovo, but also government departments and central SOE groups such as the Ministry of Public Security, the State Taxation Administration, CITIC Bank, China Merchants Bank, and China Resources Group.

 

Challenge

The Face++ product line includes dynamic facial recognition, text recognition, offline live detection, document recognition, pedestrian detection, trajectory analysis, etc. Its core applications are inseparable from deep learning technology, which trains neural networks through deep learning to improve the accuracy of face recognition.

 

Deep learning is a branch of machine learning, which uses algorithms to enable machines to learn patterns from large amounts of historical data, thus enabling intelligent recognition of new samples or predictions of the future. Deep learning first uses unsupervised learning to pre-train each layer of the network; each time unsupervised learning only trains one layer, and uses the training results as the input for the higher layer; finally, supervised learning is used to adjust all layers. In deep learning, each image has several billion or even tens of billions of connections to be processed, and training such a large network requires trillions of floating-point operations. The traditional computer system architecture cannot accelerate processing speed on a large scale. How to process data quickly and shorten the training process has become a pain point for users.

 

Solution

After analyzing the requirements, Jinpin recommended the KG4204-G5 AI Server deep learning computing platform, which is equipped with a solution of five NVIDIA Tesla V100 computing accelerators, and accelerates computing speed through the parallelism of GPU multi-cores to meet the demand for deep learning mass training data processing.

 

Jinpin KG4204-G5 AI Work Station

Features:

Supports 5 Nvidia Tesla GPU accelerators

Supports 2 Intel Xeon Scalable Platinum processors

Up to 560 TFLOPS of deep learning performance

High reliability and availability

 

Result:

The rapid rise of deep learning is due to the application of GPUs, whose powerful parallel computing capabilities greatly accelerate the speed of deep neural network training. Face++ uses NVIDIA GPU accelerators to reduce computation time by tens to hundreds of times and has achieved world-class results in the AVA and WAD open competitions.

 

The Jinpin KG4204-G5 AI server plays a critical role in the training of deep neural networks, improving the efficiency and iteration speed of developers and playing a crucial role in improving the accuracy of face recognition and target detection. By deploying the Jinpin KG4204-G5 server on a large scale, Face++ completed various complex facial recognition training in a very short time and became a unicorn in the facial recognition industry.