Table of contents
1. What is face recognition and its application scenarios
Face recognition specifically refers to computer technology that uses analysis and comparison of facial visual feature information for identity identification. Face recognition technology is mainly used for identity verification. Common scenarios include communities, buildings, campuses, factories, parks, banks, etc., such as: smart access control, face recognition gates, face attendance, smart door locks, etc. Through face recognition to verify identity, the safety of relevant places is guaranteed, and the cost of manual review is also reduced.
In security monitoring, face recognition is also of great significance, such as the security deployment of public places (subway stations, stations, streets, hotels, etc.), and the pursuit of suspects by the Ministry of Public Security. Based on the security surveillance cameras in public places, by capturing faces and uploading the results to the Ministry of Public Security network, they are compared with the faces of suspects to assist public security officers in their law enforcement work.
Face recognition technology is a key technology in the field of artificial intelligence, and has a very broad application prospect in intelligent video surveillance systems.
2. Composition of face recognition
2.1 Front-end image acquisition
The front-end face image acquisition system collects face images, real-time video streams, etc. through snapshots, such as video surveillance cameras, smart attendance machines, smart access control machines and other equipment, all have the function of face capture and snapshots.
2.2 Back-end intelligent platform
The back-end intelligent platform can gather, process, store, apply, manage and share the relevant data collected by the front-end in a unified way, and combine with the face recognition system to realize the face recognition function, and apply it in real scenarios such as: face recognition attendance , face access control, face tracking by the Ministry of Public Security, arresting suspects, etc.
3. The process of face technology
3.1 Face detection
Face detection is mainly used for preprocessing of face recognition in practice, that is, to accurately mark the position and size of the face in the image.
Find the position of the face from the photo, take the upper left corner of the picture as the coordinate origin, record the coordinates of the upper left corner and the lower right corner of the face frame, and cut out the face part.
3.2 Face Alignment
In real-world scenarios, the faces captured by the front-end equipment are often not necessarily at the frontal angle, so the face posture in the image needs to be corrected. The key point coordinates of the face are obtained through face key point detection, and the angle of the face is adjusted according to the key point coordinates of the face to align the faces. As shown in the figure, these two faces are completely different faces in the eyes of the computer, so we need to align the faces through some affine transformations.
Face alignment can greatly improve the accuracy of face recognition, because the key points of the face can make a space normalization of the face to be recognized. The operation of this space normalization is called face alignment. This operation can The subsequent extraction has nothing to do with the position of the facial features, but only features related to the shape and texture of the facial features.
Face alignment is to align the eyes, nose, and mouth of the face to a preset fixed position.
3.2.1 Affine Transformation
The function of affine transformation is a linear transformation from two-dimensional coordinates to two-dimensional coordinates, and maintains the "straightness" and "parallelism" of two-dimensional graphics (the relative positional relationship between straight lines remains unchanged, and parallel lines After the affine transformation, it is still a parallel line, and the position order of the points on the line will not change)
3.2.2 Alignment method
Use the trained model to automatically mark 68 feature points (landmarks) from the detected face, then look for a standard template in the template library, and use affine change to combine the 68 points with the 68 of the template point alignment.
There are 68 key points,
6 key points
, and 81 and 106 key points. Questyle claims to be able to extract 1000 key points.
3.3 Face encoding (extracting feature vectors)
A model is trained through a convolutional neural network, and the input model face picture is automatically encoded into a 128-dimensional vector with a strong semantic meaning.
3.4 Face Classification
Calculate the Euclidean distance between the input image and the 128-bit vector difference of each picture in the database in turn until the one smaller than our threshold is found. At this point, the face recognition is successful.
