In-depth analysis of the algorithm of face recognition technology

    Face recognition algorithms have been developed to this day and can be roughly divided into two categories: feature-based face recognition algorithms and appearance-based face recognition algorithms. Among them, most feature-based face recognition algorithms belong to early face recognition algorithms and are no longer used. However, in recent years, some new feature-based algorithms have emerged and achieved good results. Appearance-based face recognition algorithms have received widespread attention due to their simple implementation. The following will introduce the two types of face recognition algorithms respectively.

    Feature-based face recognition algorithms: Early face recognition algorithms were mainly based on feature templates and geometric constraints. This type of algorithm first processes the input image to extract facial features and appearance contours such as eyes, nose and mouth. Then the geometric relationships between these facial features, such as distance, area and angle, are calculated. After the input image is converted into a geometric feature vector, standard statistical pattern recognition techniques are used for matching and classification. Since the algorithm uses some intuitive features, the amount of calculation is small. However, since the required feature points cannot be accurately selected, its application range is limited. In addition, when the illumination changes, the face is blocked by external objects, and the facial expression changes, the features change greatly. Therefore, this type of algorithm is only suitable for rough recognition of face images and cannot be applied in practice. All
　　
    of the above methods detect specific facial features through some feature templates and geometric constraints, and calculate the relationship between the features. Some other methods use local representations of the image to extract features. Among them, the most popular method is the local binary pattern (LBP) algorithm. The LBP method first divides the image into several regions, thresholds the pixels in the 3x3 neighborhood of each region with the center value, and regards the result as a binary number. The characteristic of the LBP operator is that it remains unchanged for monotonic grayscale changes. Each region obtains a set of histograms through such operations, and then all the histograms are connected to form a large histogram and histogram matching calculations are performed for classification.
　　
    The main advantage of feature-based face recognition algorithms is that they are robust to changes in posture, scale, and illumination. Since most features are based on manual selection and prior knowledge, they are less affected by the imaging quality of the image itself. In addition, the extracted facial features are often of low dimension and have a fast matching speed. The disadvantage of these methods is that automatic feature extraction is difficult. If the discriminative ability of the feature set is weak, no amount of subsequent processing can compensate for its own shortcomings. Appearance-based face recognition algorithms: Appearance-based face recognition algorithms are also called holistic methods. They use the global information of the image to identify faces. The simplest holistic method is to use a two-dimensional array to store the grayscale values ​​of the image, and then directly compare the correlation between the input image and all images in the database. This method has many disadvantages, such as being easily affected by the environment and time-consuming to calculate. One of the important problems is that such classification is performed in a very high-dimensional space. In order to overcome the dimensionality problem, some algorithms use statistical dimensionality reduction methods to obtain and retain more useful information. The most typical algorithms are principal component analysis (PCA) and linear discriminant analysis (LDA). The
　　
    PCA algorithm points out that any specific face can be represented by a low-dimensional feature subspace and can be approximately reconstructed using this feature subspace. The features obtained by projecting the input face image onto the feature subspace are compared with the known database to determine the identity. The features selected by the PCA algorithm maximize the differences between face samples, but also retain some unnecessary changes caused by lighting and facial expressions. The changes caused by lighting for the same person may be greater than the changes between different people. The LDA algorithm minimizes the sample differences within the same individual while maximizing the sample differences between different individuals. This achieves the division of the face feature subspace. The main advantage of holistic methods is that they do not discard any information in the image. However, this is also their disadvantage. Holistic methods generally assume that all pixels in the image are equally important. Therefore, these techniques are not only computationally time-consuming, but also require that the test samples are highly correlated with the training samples. Face recognition performance is average when the posture, scale and lighting of the face image change significantly.