Brand New Portrait Recognition Technology
The key to a successful portrait recognition system lies in having cutting-edge core algorithms that yield practical recognition rates and speeds. It integrates various specialized technologies such as artificial intelligence, machine learning, model theory, and video image processing. How to improve the algorithm so that the portrait recognition technology can accurately identify faces while adapting to the changing environments and characteristics of personnel in real-world scenarios, and at the same time meet the conditions of being easy to promote and cost-effective, has become the breakthrough point for the next generation of portrait recognition technology.
The Sino-German Hongtai portrait recognition system is based on industry-leading artificial intelligence analysis algorithms, providing a complete set of technical solutions including face detection, face recognition, key point positioning, body appearance identification, and gait feature analysis to address the technical challenges of dynamic portrait recognition that are easily influenced by environmental factors.
Traditional portrait detection and recognition algorithms follow a fixed phased training framework, where manually defined feature representations are often too simplistic to adequately represent portrait features. The segmented training approach means that the objectives of each training segment are not entirely consistent, which restricts further improvements to the algorithm. In contrast, the complex nonlinear feature representation and “end-to-end” nature of deep learning are entirely different; the entire training process is directed by a unified goal, with no need for manual definitions in the middle process, allowing for the learning of more representative feature representations, thus achieving better results. Deep Vision continuously integrates gait and body appearance features to further enhance recognition capabilities in real-world scenarios.
The Sino-German Hongtai portrait recognition system, Deep Vision, uses deep learning neural network algorithms. Deep learning mimics a cognitive structure where the lower-level features are progressively abstracted into higher-level features, which are then expressed well with images of these features. Deep Vision has been deeply optimized for various complex scenarios, significantly improving the portrait recognition rate in complex environments while effectively controlling the false positive rate. Specific optimizations have been made for the deep network structure and training datasets to improve computational speed while ensuring recognition rates. Combined with distributed computing systems, it can achieve large-scale, highly available practical systems.
Modular Design for Rapid System Installation
The Sino-German Hongtai portrait recognition system adopts a flexible modular design that supports the flexible assembly of systems to meet different project requirements for portrait recognition. Its settings and operations are quite simple and easy to use, allowing for quick learning and management. The system installed on servers with different configurations will have varying numbers of access channels. The Deep Vision portrait recognition system tested this time was installed on an X86 high-performance server (i7 processor, 16GB RAM, running 64-bit Windows 7 operating system) and matched with a 1080P HD camera with a resolution of 1050TVL, with both the server and the camera connected to a switch.
The system uses SQL Server 2008 as its database and includes five major server modules: face capture server, face library/casual face library server, waiting-to-compare face library server, face comparison server, and business server. The portrait recognition system accesses these five servers through a comprehensive control client. Furthermore, the face capture server can directly perform key point positioning, body appearance feature identification, and gait feature analysis, effectively distinguishing features between different individuals and identifying personnel entering and exiting. It is worth mentioning that this portrait recognition system can be used as an independent project or in combination with other platforms or projects, providing great flexibility.
User-Friendly Client
The Deep Vision portrait recognition system control client features a simple design with four main categories: System Configuration, Real-Time Monitoring, Face Library, and Query. The UI design is minimalistic and user-friendly, making it easy to learn, operate, and manage.
The system configuration includes local configuration, front-end configuration, service configuration, and user configuration. Local configuration involves alarm prompts, image capture, and capture path settings; front-end configuration is for adding cameras to obtain real-time video streams. This recognition system supports third-party cameras and mainstream brand cameras from home and abroad that comply with ONVIF, PSIA, and RTSP protocols, and it can connect to analog, standard definition, high definition, and ultra-high definition video sources; service configuration is for adding capture servers and face comparison servers to the local client to perform detection, comparison, and recognition operations of faces in the video stream; user configuration is for managing user additions and deletions.
Face Library
The face library includes creating a face library, browsing the face library, casual face library (faces captured that are not in the comparison database), and batch import.
When creating a face library, the recognition threshold can be set. First, a new face library can directly obtain connected cameras and move desired cameras into the face library list. The system supports both single image data uploads and batch import functions. When uploading a single image, personal information such as name, gender, birthday, document type, and document number can be directly entered, followed by uploading the face image. The face library has an automatic face detection function, capable of automatically reading faces from any image of unlimited size in JPG, PNG, etc., without the need for special processing and optimization of the images; face import is very convenient, and a single database supports importing multiple face images. During batch import, personal information can be modified separately in the database, allowing for quick establishment of the database. Compared to importing images one by one, this saves time. In this evaluation, face images captured from cameras, DSLR photos, and online images were extracted and uploaded to the database in JPG format. It should be noted that the images uploaded to the face library in this evaluation contained multiple faces per image, which were of lower quality and pixel count compared to standard face images, but they could still be recognized, and the face database was created quickly.
Browsing the face library allows for viewing all created face libraries, with browsing results displayed in the order of upload, with each item containing all necessary upload items.
The casual face library can query all faces captured by cameras, which are all face resources of casual individuals. Faces can also be exported and compared with the blacklist library.
Comparison Query
The query is divided into face comparison queries and comparison log queries.
In the face comparison query, only one photo needs to be uploaded, and the query library (which can be a face library, comparison log, or casual face library) and camera can be selected. The query results are displayed in order of similarity from high to low, and results can be exported and saved. In practical tests, uploading a frontal photo of a face already existing in the database yielded similarity results ranging from 69.3% to 78.6%, with lower similarity results also displayed.
The comparison log query allows for searching by name, ID card number, and gender, with query results displaying the desired individual’s information (such as location and time of appearance).
Real-Time Portrait Recognition
The Sino-German Hongtai Deep Vision portrait recognition system can monitor multiple cameras simultaneously, with interface settings divided into one-screen, four-screens, and nine-screens. In the real-time monitoring screen, detected portraits can automatically be compared with the face library, triggering alarms when similarity exceeds the set threshold, and displaying the number of successfully matched similar image data on the real-time monitoring interface.
The system has strong anti-interference capabilities, allowing for recognition under conditions such as wearing hats, hairstyles, makeup, exaggerated expressions, and large angle tilts. It can analyze multiple individuals and operate in all-weather black and white/full color modes, achieving impressive portrait recognition results and demonstrating strong practical capabilities. The following types of portrait images were tested and recognized:
Normal Portrait Recognition: The camera was aimed at the test scene, with the tester facing the camera; the real-time monitoring interface immediately displayed the tester’s portrait, and the system automatically compared it with the tester’s face library and triggered an alarm upon successful comparison (the face library had the tester’s photo uploaded).
Low Pixel Portrait Recognition: In the test scene, when the tester was positioned differently and the depth of field was large, the displayed portrait had a low pixel count, with the minimum recognizable portrait being 60×60 pixels, and the recognition time was visually less than 1 second, completing the comparison quickly.
Ultra High Pixel Portrait Recognition: Supports recognition of maximum image sizes of 8192×4096.
Profile Recognition: Supports recognition of faces at angles. In tests, the maximum pitch angle was 30°, with head tilts of 20°-30° and horizontal flips (tested with photos), and the system accurately recognized them. When the rotation angle reached 45°, the system could detect the face, but due to limited recognizable features, the recognition rate against the comparison database was not high.
Multi-Person Recognition: In a test scene containing multiple portraits, the maximum number of recognizable portraits in one frame was 20.
Complex Feature Expression Recognition: During tests, whether smiling, crying, closing eyes, rolling eyes, or making funny faces, the system achieved over 90% recognition rate.
Abnormal Face Recognition: In tests, when one-third of the left or right side of the face was blocked, wearing a hat, or mimicking wearing a mask covering the nose and lower face, the system could still recognize the face and compare it with the face library.
Complex Scene Portrait Recognition: Initially, single and multi-person recognition was tested against complex backgrounds, and the system could accurately capture and recognize portraits. Next, low-light scenarios were simulated; the paired 1080P camera provided excellent images, enabling normal recognition even in low light. When simulating strong light scenes, recognition results were still acceptable; however, in simulated scenes, the recognition rate decreased.
Diverse Application Scenarios
The Sino-German Hongtai Deep Vision portrait recognition system addresses the shortcomings of two-dimensional face recognition, which can lead to inaccuracies and missed detections, by introducing a brand new 3D facial biometric recognition that can automatically perform three-dimensional modeling analysis of faces, supporting pitch angles of 30° and lateral angles of 30° for facial feature analysis. It also incorporates various algorithmic anti-interference engine technologies to eliminate certain face interference factors, such as glasses, hats, beards, makeup, and even masks. It adapts to multiple person recognition, scene changes, variations in environmental illumination, and complex adverse conditions such as strong light/glare. To effectively implement portrait recognition from the laboratory to practical applications, with strong adaptability to real-world scenarios and improved practical application effectiveness, Sino-German Hongtai has also designed targeted features for the system, such as dynamic collection technology, which only captures images when the “portrait” in the frame is in motion, to avoid interference from wall paintings or posters that may occur in real scenarios. This achieves dynamic real-time recognition with high recognition rates and speeds reaching the second level.
Product Features
1. Modular design, flexible selection and assembly to meet different project portrait recognition needs;
2. Simple interface setup, easy to use, and convenient management;
3. Strong anti-interference performance, capable of recognizing under conditions such as wearing hats, hairstyles, makeup, exaggerated expressions, and large angle tilts;
4. Adaptable to portrait recognition in complex environments (complex backgrounds, many personnel, large age time spans, low illumination, etc.);
5. 3D facial biometric recognition technology and dynamic collection technology effectively avoid interference from wall paintings, posters, and similar distractions in practical application scenarios.
Testing Locations and Supporting Equipment
a&s testing room, high-performance computer with internet access, switch, DSLR camera, Skyworth 6M48 monitor, various brands of 1080P/720P cameras, multiple testers, photo/artistic images, etc.
Applicable Scenarios
Suitable for applications in roads, airports, communities, train stations, shopping malls, and other industry applications, as well as civil applications such as company attendance and security deployments.
Specifications
Maximum face recognition size: 8192×4096, minimum: 60×60 pixels
Face recognition angles: pitch 30°, lateral 30°
Maximum number of faces recognizable simultaneously: 20
Anti-interference factors: hats, hairstyles, makeup, exaggerated expressions, large angle tilts
Frontal recognition rate: over 98%
Anti-interference recognition rate: over 75%
《2015-2020 System Integration Industry Development Trends and Analysis》
↓↓ What You Want ↓↓