Development of Face Recognition_ Taigewang Technology

The classic process of automatic face recognition is divided into three steps: face detection, face feature point location (also known as face alignment), feature extraction and classifier design. Generally speaking, face recognition in a narrow sense refers to & quot; Feature extraction classifier & quot; Two parts of algorithm research. Before the emergence of deep learning, face recognition methods were generally divided into two steps: high-dimensional artificial feature extraction (such as LBP, Gabor, etc.) and dimension reduction. Representative dimension reduction methods include subspace learning methods such as PCA and LDA and popular learning methods such as LPP. After the popularity of deep learning methods, the representative method is to directly learn the discriminant face representation from the original image space. Generally speaking, the research history of face recognition can be divided into three stages. In the first stage (1950s-1980s), face recognition was regarded as a general pattern recognition problem. The mainstream technology was based on the geometric structure of the face. In the second stage (1990s), face recognition developed rapidly, and many classical methods appeared, such as eigenface, Fisher face and elastic graph matching. At this time, the mainstream technical route is face apparent modeling. In the third stage (from the end of the 1990s to the present), the research on face recognition has been deepening. Researchers began to pay attention to face recognition facing real conditions, mainly including the following four aspects: 1) propose different face space models, including linear modeling methods represented by linear discriminant analysis, Nonlinear modeling method represented by kernel method and 3D face recognition method based on 3D information. 2) Deeply analyze and study the factors affecting face recognition, including illumination invariant face recognition, pose invariant face recognition and expression invariant face recognition. 3) New feature representations are used, including local descriptors (Gabor face, LBP face, etc.) and deep learning methods. 4) Using new data sources, such as face recognition based on video and face recognition based on sketch and near infrared images. Since 2007, LFW database has become the test benchmark for face recognition under real conditions. The LFW dataset includes 13233 face images of 5749 people from the Internet, of which 1680 people have two or more images. The standard test protocol of LFW includes the ten fold confirmation task of 6000 pairs of faces. Each fold includes 300 pairs of positive examples and 300 pairs of negative examples. The ten fold average accuracy is used as the performance evaluation index. Since the release of LFW, the performance has been constantly refreshed. Before 2013, the main technical route was artificial or learning based local descriptor measure learning. After 2014, the main technical route is in-depth learning. Since 2014, deep learning big data (massive labeled face data) has become the mainstream technical route in the field of face recognition. Two important trends are: 1) the network becomes larger and deeper (vggface16 layer, facenet22 layer). 2) With the increasing amount of data (deepface 4 million, facenet 200 million), big data has become the key to improve the performance of face recognition. In the pre DL era, taking the third-generation and half SDK of vipl laboratory as an example, the key technical points include 1) block face feature fusion: Gabor feature LPQ feature. 2) Subspace learning for feature reduction (PCA LDA). 3) Fusion of multi-scale face normalization template. The related technologies of sdk3.5 achieved a confirmation rate of 96% under the condition of 0.1% error acceptance rate in FRGC Experiment 4, which is still the best result in FRGC data set. It should be noted that although deep learning emphasizes feature learning, learning features is not the patent of DL. In the pre DL era, the work of directly learning representation from images using shallow models and learning semantic representation based on artificial descriptors (such as attributes and simile classifier for learning middle-level attribute representation and Tom vs Pete for learning high-level semantic representation) are seen in relevant literature. In 2014, Facebook published its work on cvpr14. Deepface combined big data (4 million face data) with deep convolution network, which approached human recognition accuracy on LFW data set. Deepface also introduces a local connected convolution structure to learn a separate convolution kernel at each spatial position. The disadvantage is that it will lead to parameter expansion. This structure did not become popular later. Deepid family can be regarded as a group of representative work in the field of face recognition in the DL era. The earliest deepid network included four convolution layers and adopted softmax loss function. Deepid2 considers both identity loss and verification loss on the basis of deepid2 network. These two losses can be realized by using softmaxwithlos layer and contrast loss layer respectively in Caffe deep learning framework. Deepid2 network adds auxiliary loss function of each layer on the basis of deepid2 (similar to deep supervised network). Google published its work in cvpr2015. Facenet uses 22 layers of deep convolution network and massive face data (200 million images of 8 million people) And the triple loss function commonly used in image retrieval tasks. It is worth mentioning that, since the number of face categories reaches 8 million, if softmax loss is used, the number of output layer nodes will reach 8 million, and at least 32GB video memory is required (assuming that the last hidden layer node has 1024, using single precision floating-point numbers) However, triplet loss does not need to occupy additional video memory. Facenet's average accuracy of 10% discount on LFW data set reaches 99.63%, which is also the best result in the officially published papers so far, almost announcing the end of the eight-year performance competition on LFW from 2008 to 2015.