Face recognition lighting adaptation in monitoring applications

As a public security organ at the public security and prevention and control field, according to need, a video surveillance network that basically covers complex social security areas and important sites has been established. Although the number of cameras of different categories is increasing rapidly, these cameras are in advance precautions. Still did not play enough role. For example, face images acquired using existing video surveillance systems may not meet the needs of face recognition-based identity authentication technology due to low resolution and uneven illumination. Therefore, the construction of a smart monitoring system based on face recognition can be widely applied not only in the tracing of criminal suspects, identity recognition, exit and entry management, important places or system authentication, but also to monitor the flow of people in key locations and densely populated areas. There are also great potential applications.

However, as a face recognition application unit, face recognition has many interference factors in the specific application process, which not only affects the face detection (face image preprocessing) directly affects the accuracy of face recognition. Especially the influence of natural light on the system is particularly prominent. To this end, we invited relevant experts and technicians to conduct research on video surveillance scenes that can use face recognition technology, and at the same time make experimental adjustments on some cameras. Aiming at the problem of face illumination, a multi-scale and multi-directional face edge feature extraction method based on dual-tree complex wavelet was proposed. The dual-tree complex wavelet denoising model was used to extract the illumination-invariant face features and face features. In combination with the edge information, an enhanced face feature map is constructed as the invariant feature of the face illumination. In face-invariant feature extraction under illumination conditions, a non-aligned robust face-based super-resolution method based on iterative sparse representation is effectively integrated to overcome the problem of face image super-resolution in the case of misalignment.

For the fuzzy problem of the face image of video surveillance system, we have targeted the optimization of the recognition algorithm to extract the anti-fuzzy features that are compact and have strong description ability from the blurred face images. This feature still has better recognition accuracy for blurred images on real data with more complex environment. This method also has obvious advantages on the accuracy of fuzzy image recognition.

Through continuous exploration and practice, and multiple trials in a variety of scenarios, according to actual use and data analysis found that can effectively solve the captured images of dynamic video images, two eyes are too small pixels, dynamic blur, light discontinuity and backlighting / Interference caused by unfavorable factors such as side light. Without changing the image quality evaluation criteria, the number of images that can be effectively used for face recognition increased from 46.8% to 61.9%. Using second-generation ID photos as database templates, the library capacity is 10,000. Compared with the original system, the average false positive rate has dropped from 1.72% to 0.75%, while the average accuracy rate has increased by 5.4%.

Through the above-mentioned practice, it is demonstrated that the technology has a significant role in improving the accuracy of face recognition in the video surveillance environment under conditions of low resolution or dramatic changes in light illumination, and can significantly improve the face recognition system for uncontrollable video surveillance environments. Adaptability provides a strong technical support for face recognition technology to provide deeper and more comprehensive services to police officers in the public security industry.