Face Detection and Recognition
The ANPR software detects number plates within a given image and reads a given text. The system can be completed by a car category/make/model recognition module and by a face detection module.
- Fast face detector robust to illumination changes
- Gender and age recognition module
- Person ID tracking
- Identity recognition module
- Audience measurement systems
- People counting systems
- Face recognition systems
- Human machine interfaces
A standalone application or SDK.
The audience measurement system developed for Media Research a.s.
How does it work?
Our team has been developing face detection and recognition methods for more than 10 years. Currently we are using detectors based on sequential classifications that allow for extremely fast data evaluation.We construct our sequential detectors using the automated learning method based on WaldBoost algorithm. This algorithm, on the basis of a wide series of images (500 000 positive examples, 4*10^9 negative examples) finds the optimal decision rule, which maximises both the detection success and the speed of evaluation. Because of this unique process our detectors are extremely fast and precise.
Effectiveness comparison versus OpenCV detectorsFor illustration purposes we have included two graphs (see Fig. 1 and Fig. 2) which compare the effectiveness of our detector compared to detectors from ‘open source’ library OpenCV. The test was conducted on two independent test series. The first graph measures the precision results of a standard test series CMU-MIT, the second on our internal series. It is clear from both tests that our detectors are up to 10% more precise.
Speed comparison with Open CV detectorsIn comparison with Open CV detectors our routines are more than 10 times as fast. Table 1 shows the results of the speed of a sequence scanned in standard DV resolution of 720x576 pixels. The speed was measured on one core of ordinary PC with a 2.8GHz CPU.
|Table 1: Speed comparison of sequence detection with a resolution of 720x576 pxl|
|Speed||9.26 fps||1.42 fps||1.83 fps|
- Software routines for face detection are supplied in the form of SDK comprising of a dynamic library (dll, so), a simple example in C/C++, demo application and documentation
- Others as per requirements
Custom made face detectorsEyedea Recognition also develops custom-made face detectors that are optimised for customers given conditions and requirements. For example: detection in panorama images, detection in very low resolution etc.
Face RecognitionEyedea recognition develops and supplies software routines for face detection and recognition in a camera shot. We mainly concentrate on:
- Gender recognition,
- Age recognition,
- Identity recognition.
Gender RecognitionData required for gender recognition is a frontal face view with a minimum of 40x60 pixels resolution. The result is a gender recognition assessment with a classification error being in the vicinity of 5% - 10% depending on the difficulty of the test set.
Age recognitionData required the age recognition is also a frontal face view with a minimum of 40x60 pixels resolution. The result is an estimate of age in years.
Identity recognitionThe module for face recognition can recognize up to 100 different faces in a camera shot. The data required is once again a frontal face view with a minimum of 40x60 pixels resolution. The result is identity recognition.
- Software routines for face detection and recognition are supplied in the following forms
- Standalone application, or
- SDK comprising of a dynamic library (dll.so), a simple instruction demo in C/C++, demo application and documentation
- Others as per requirements