In the whole system of intelligent transportation, license plate recognition is undoubtedly one of the basic applications. The application of license plate recognition requires the extraction and recognition of the stationary or moving vehicle license plate from the complex background. Through license plate extraction, image preprocessing, feature extraction, license plate character recognition and other technologies, the identification of vehicle license plate, color and other information. License plate recognition is the basic application of automatic traffic control, and its recognition success rate and accuracy rate will have a great impact on traffic operation efficiency, toll collection, punishment for illegal behaviors, etc.
Beijing Zhixin Original Motion Technology Co., LTD. *(hereinafter referred to as: Smartcore is a leading provider of artificial intelligence and solutions in China, specializing in the technology research and development of artificial intelligence algorithms and algorithm chips, the development and delivery of intelligent products and solutions, as well as the integration of intelligent cloud services, committed to accelerating the application of artificial intelligence technology in various industries, in order to improve user experience and work efficiency. In order to meet the requirements of license plate recognition, Zhixin Original Motion launched the Shenyun Zhiche * platform, which can accurately recognize all kinds of license plates. In addition, Zhixin has also launched overseas license plate recognition solutions, which can be iterated quickly under the condition of a small number of (>1K) license plate samples, and only 2-4 weeks of delivery cycle to achieve the development task of new national license plates, and the comprehensive accuracy rate can be as high as 90% to 95%.
In order to further improve the reasoning performance of the license plate recognition platform and accelerate the algorithm training speed for overseas license plate scenarios, the intelligent Core is equipped with the combination solution of the second generation Intel ® Xeon ® scalable processor and Intel ® OpenVINO™, and uses the Caffe* application optimized for Intel ® architecture. The performance has been improved by tens of times.
Challenges: Overseas markets pose a serious test for license plate recognition
In the field of Internet parking, in addition to the license plate recognition at the entrance and exit, Smartchip has also launched roadside parking cameras and parking cameras in the parking lot. At the same time, Smartchip has also begun to enter the overseas parking market, and has been deployed in Hong Kong, Taiwan, Macao, Malaysia, Indonesia, Thailand, Singapore, Vietnam and other places. It is expected that the number of license plate recognition cameras deployed in the next 3 years is close to one million. In the overseas license plate recognition market, an important feature is that each country or region has certain differences in the length, width, color and character composition of the license plate. The same algorithm cannot be applied to every country and region.
In order to train the license plate recognition algorithm of each country and region, there is no doubt that localized data is needed as support. With traditional license plate algorithms, it would take a huge number of license plate samples and a months-long delivery cycle to improve the accuracy of license plate recognition to a usable level.
In addition, license plate recognition system itself also brings requirements for deep learning performance. According to the calculation of the engineers of Zhixin Original movement, it takes 40ms for a single core processor to recognize a car model at present. According to the calculation of one car every 10 seconds during the parking peak, the deep cloud car recognition service needs a total of 1000 4-core processors. Once the police service is provided, the pressure will increase exponentially, so it is necessary to realize more efficient deep learning computing capabilities.
Smart Core is the leading algorithm platform for overseas market
To meet the demand of overseas market for license plate recognition, Smart Chip has launched overseas license plate recognition solution. Technically, the realization of overseas license plate algorithm is completed through three steps of license plate extraction, character segmentation and character recognition, that is, extract license plate image from the image containing license plate with complex background, and then carry out necessary preprocessing to extract the image and separate out a single character. Then the character features are extracted and compared with standard characters, and the license plate number of the license plate to be recognized is output. Zhixin original motion for different license plate types will use different license plate location and character segmentation algorithm to ensure the accuracy of recognition.
The biggest highlight of the scheme is that it is based on the self-developed license plate algorithm framework, which can quickly iterate under the condition of a small number of license plate samples (>1K). The development task of new national license plates can be realized in a delivery cycle of only 2-4 weeks, and the comprehensive accuracy rate can be as high as 90%-95%. At present, the program has been commercialized in more than 20 countries and regions. In addition, the programme has the following features:
● Support more than 20 countries and regions license plate recognition
● Applicable to bayonet, entrance, parking space and other scenes
● The algorithm framework adopts flexible model matching strategy and modular design, which can realize the development and implementation of new national license plates under a small number of samples
● The algorithm supports cross-platform design to meet the requirements of front-end camera and back-end server
● China team original research and development support, problem iteration efficient, timely response to project needs
In this scheme, the intelligent core primitive motor uses the deep learning method based on convolutional neural network for vehicle model recognition, and realizes the classification reasoning optimization through the topological structures such as MobileNet* and GoogleNet*. After optimizing the license plate recognition algorithm, it is helpful to realize the rapid development and deployment of applications on the premise of small samples.
Solution: Accelerated reasoning performance based on Intel ® architecture
In terms of infrastructure architecture, Smartcore original motion has launched different reference schemes for the actual application environments of different countries and regions. The workload of license plate recognition can be flexibly carried by cloud data centers or edge devices, which can meet the different requirements of users on delay, deployment cost, network and other aspects.
On the server side, SmartCore is equipped with a second generation Intel ® Xeon ® scalable processor and Intel ® OpenVINO™ solution. The second generation Intel ® Xeon ® scalable processor helps enable full AI support from the entire data center to the edge.
Compared to the first generation Intel ® Xeon ® expandable processor used before the core, the second generation Intel ® Xeon ® expandable processor further improves the performance, especially the support of VNNI and other technologies to improve the reasoning performance to a new level. In specific application cases, performance improvements will allow users to deploy fewer nodes while supporting more reasoning load and achieving lower total cost of ownership (TCO).
Intel ® OpenVINO™ Tool Suite Distribution enables developers to easily integrate deep learning reasoning into applications using industry-standard AI frameworks, standards, or custom layers. By running on an underlying Intel ® Xeon ® scalable processor, the Intel ® OpenVINO™ Tool Suite distribution achieves competitive reasoning speeds and minimal loss of accuracy. With AVX-512 and MKL/ McL-dnn boost libraries, the solution also delivers superior computing performance.
Effect: A nearly 30-fold performance improvement
In order to verify the reasoning performance of the second generation Intel ® Xeon ® scalable processor and Intel ® OpenVINO™ in different topologies, a test platform was built for the Intelligent Core. The configuration of the test platform is shown in Table 1.
In the reasoning performance test of video image analysis, the reasoning performance of the public Caffe*, the Intel optimized Caffe, and OpenVINO™ was tested respectively in several topologies such as MobileNet, Mobilenet-V2, GoogleNet, and VGG-16. The test data, shown in Figure 1, showed a 28.4-fold performance improvement in MobileNet using OpenVINO™ compared to Intel optimized version Caffe.
Meanwhile, in the test of SSD and RPN detection inference optimization (test data is shown in Figure 2), compared with Intel optimization Caffe with FP32, Intel optimized Caffe with SSD-VGG and RPN-VGG topologies and Int8 quantization can achieve 2.58x and 2.09x performance improvements, respectively.
Future: Provide flexible and efficient computing support for vehicle identification application loads
In addition to license plate recognition, detection reasoning based on artificial intelligence and deep learning technology is also widely used in other scenarios of vehicle recognition, and used to detect vehicle model, color, size, location and other purposes. These application loads not only rely on advanced algorithms, but also put forward certain requirements for the computing power of the platform. The cloud vehicle model identification platform of Zhixin Original Movement can identify about 1,600 vehicle models, including brand, model and age, with an identification accuracy of more than 99%. The platform can also be applied to the identification of vehicle models in the police security system, highway toll collection and other fields.
Through cooperation with Intel, Smartcore Primitive can provide a more flexible infrastructure platform for deep learning applications of vehicle recognition. For example, deep learning heterogeneous computing can be realized through the combination of CPU, GPU and FPGA in the data center, or unified edge computing server can be deployed at the edge end to carry the load of vehicle and license plate recognition. Meet the needs of users for detection accuracy, speed and other aspects, and help the realization of intelligent transportation.
1. Based on the internal test results of Intel and Smartcore Original motion
2. Test configuration: Platform 1 -- Intel ® Xeon ® Gold 6140 processor; BIOS version (including micro code version: the cat/proc/cpuinfo | grep microcode - m1) : SE5C620. 86 b. 00.01.0012.021320180053, the total memory 96 GB (six slots / 16 GB / 2666 MHZ). One Intel 960GB solid State disk operating system disk; CentOS 7.4.1708, 3.10.0-957.10.1.el7.x86_64; Platform 2 - Intel ® Xeon ® Platinum 8260 Processor; BIOS version (including micro code version: the cat/proc/cpuinfo | grep microcode - m1) : SE5C620. 86 b. 02.01.0008.031920191559, the total memory 96 GB (six slots / 16 GB / 2666 MHZ). One Intel 960GB solid State disk operating system disk; Ubuntu 16.04.3 LTS, 4.10.0-28-generic
