Although face recognition technology has gone through a long research stage, it is still considered to be one of the more difficult research topics in biometric recognition technology, for the following reasons:

1. The background environment is complex and diverse

Before face recognition, it is necessary to locate the face in the monitoring scene, that is, face detection.Face detection is correct or not directly affects the performance of face recognition.When the background of the monitoring scene is more complex, the face detection rate will decrease accordingly. Therefore, the face detection algorithm that can adapt to the complex background environment is one of the difficulties of face recognition technology.

2. Complex and changeable lighting conditions

In the practical application of intelligent video surveillance system, the detected face images will have different dark changes due to the change of monitoring environment light, as shown in Figure 1.The test of FRVT2006 shows that the performance of face recognition under different light conditions is significantly improved compared with that of FRVT2002, but the influence of light on recognition rate has not been overcome fundamentally.

                                          图:光线变化对采集到的人脸影响示意图

3. Diversity of facial expressions

In practice, the expression of a human face can change at any time.The figure below shows a partial face image with changing expressions.As can be seen from the figure below, when human expression changes, it may lead to changes in face contour and texture. Meanwhile, due to the traction of facial muscles, the position of facial feature points will also change accordingly.Different expressions cause different facial changes. In addition, different people have different effects on the same expression, so it is difficult to use a unified standard to accurately classify the effects of different expressions on different people.

                                       图:人脸表情变化多样性示意图

4. Diversity of face angles

Face Angle diversity mainly refers to the rotation of the detected face image due to different shooting angles, including plane rotation and depth rotation.Figure 3 shows some face images taken from different angles.As can be seen from the following figure, the change of shooting Angle will also lead to the change of face contour, the same as the change of expression on face image. In addition, due to the change of Angle, some features of the face may not be extracted correctly, which further leads to the wrong recognition of the face.

                                      图:采集人脸的角度多样性示意图

5. Occlusion

Even if the non-artificial deliberately occlusion, in the actual application of the detected face images often appear such as hats, glasses and other occlusion, in addition to these, the change of beard and bangs also directly affect the face feature extraction.When the face image occlusion occurs, a lot of information of the face will be lost, resulting in the face recognition algorithm error or failure.

Face recognition system mainly includes four parts: face image detection, face recognition pretreatment, face feature extraction and feature matching and recognition.

1. Face image detection

Face image detection is one of the key steps in face recognition process.Face detection refers to any given image, using a certain strategy to search it to determine whether there is a face, if there is, then return the detected face image location, size and posture.Face detection mainly uses histogram features, color features, template features, structure features and Haar features of face images.

2 face image preprocessing

Preprocessing refers to the face recognition before, in order to improve the recognition rate, through the image processing technology to detect a series of face image quality improvement.These processes mainly include gray correction, noise filtering, light compensation, histogram equalization, normalization and so on.

3. Face feature extraction

Face feature extraction is the process of face feature modeling. The extraction methods are mainly divided into two categories: knowledge-based representation methods and algebraic feature or statistical learning based representation methods.At present, face features used in face recognition technology mainly include visual features, face image transformation coefficient features, face image algebraic features and so on.Among them, the knowledge-based representation method is mainly based on the shape description of facial features and the distance between them to obtain the feature data that is helpful for face classification, and the feature component is usually using the Euclidean distance, curvature or Angle between feature points.Geometrical feature-based representation method refers to the face representation method based on the geometric description of the structural relations between the facial features.

4. Face matching and recognition

Matching and recognition refers to using the face features extracted in the previous step to search and match with the feature template stored in the sample library. In this process, a threshold needs to be defined in advance. When the similarity exceeds the threshold, the matching results will be output.

本文仅做学术分享，如有侵权，请联系删文