Transfer Learning for Face Recognition
The research history of face recognition is quite long-standing. As early as 1888 and 1910, Galton published two articles in *Nature* on using faces for personal identification, analyzing the human ability of face recognition. However, at that time, the problem of automatic face recognition was beyond reach. In recent years, face recognition research has attracted the attention of many researchers, and a variety of technical methods have emerged. In particular, since 1990, face recognition has made significant progress. Almost all well-known universities of science and engineering and major IT companies have research groups working on related studies. In the early stages, traditional face recognition was usually studied as a general pattern recognition problem. The main technical approaches adopted were geometric feature-based methods. This was largely reflected in the study of profile silhouettes, where a great deal of research was devoted to the extraction and analysis of structural features from facial silhouette curves. Subsequently, appearance-based modeling methods such as Eigenface, Fisherface, and elastic graph matching were continuously proposed. Starting from the late 1990s, researchers began to focus on face recognition under real-world conditions, proposing different face space models, including linear modeling methods represented by Linear Discriminant Analysis, nonlinear modeling methods represented by kernel-based methods, and 3D face recognition methods based on 3D information. New feature representations were proposed, including local descriptors (Gabor Face, LBP Face, etc.) and deep learning methods. Since 2014, deep learning + big data (massive labeled face data) has become the mainstream technical approach in the field of face recognition. Deep neural networks such as VGGFace, DeepFace, and FaceNet have been continuously proposed, and face recognition accuracy has been steadily improving. In 2014, Facebook's work DeepFace, published at CVPR 2014, combined big data (4 million face images) with deep convolutional networks, approaching human-level recognition accuracy on the LFW dataset. Google's work FaceNet, published at CVPR 2015, surpassed human-level recognition accuracy on the LFW dataset by adopting the Triplet Loss function.