Lei feng's network: this paper from Tencent, Lei feng's network (search for "Lei feng's network", public interest) has been authorized to do so. Introduces the three part of face recognition systems, and mobile phone cameras can automatically facial depth analysis why.
This is a "face" of the era, a few words about the human face, the most well known is face recognition. The technology in the finance, insurance, education, security and other areas of activity, become the stars in the field of AI technology. Micro letter before the public, focuses on face recognition, this paper mainly introduces some quiet support for face recognition technology. To learn more about face recognition technology can be found in the application deep learning in face recognition--grandmother model "evolution"
In General, a complete face-recognition system consists of three main components, namely, face matching for face detection and face recognition. Lines of the three actions: face detection face location found in the image, then face matching face found on the eyes, nose, mouth, facial features, such as location, finally, face recognition feature and both face than to compute similarity, confirm the identity corresponding to the face.
Figure 1 face recognition processes
1. Introduction to face registration
Registration for face (Face Alignment), also known as facial feature points detection and location. Facial feature points different from the corner or SIFT features such as image feature points in General, facial feature points are usually a set of previously defined by manual (see Figure 2). According to different application scenarios, features a different number, for example 5, 68, 82 points.
Figure 2 common facial feature points detection and positioning of target detection
In addition plays a key role in human face recognition system, face matching technology in 3D face modeling, facial animation, facial analysis, face beautify and virtual makeup, face self-timer areas such as the dynamic response of a wide range of applications. Make a small, excellent facial registration tracking performance, mainstream mobile single-frame processing speed can reach less than 3ms, already in "every day p-dynamic response of self", "mobile QQ-a short video", "mobile video chat QQ-" "mobile phone Qzone-the dynamic response of the camera" scenarios, such as landing.
Figure 3 face beautify and virtual makeup Ted Baker phone cases
2. face registration status of research on traditional face registration
And others face similar changes in illumination, head pose, facial expression and so, will greatly affect the face and occlusion accuracy of registration. Face registration but also has its own characteristics, first describes the structure of the face feature points (outline and facial features), face stability of the structure is intact, and features fixed relative position and secondly, changes in posture and facial feature points of the head position changes. Traditional face registration and to keep trying to find a more accurate characterization of expression of this assemblage of both determine and change, and then select appropriate optimization method based on descriptor, which face feature points.
Most direct feature descriptor is used in color, grayscale, and different skin for each part of the face detection location. Slightly more complex you can choose from a variety of texture features, such as based on Haar texture features and Adaboost training cascade classifier for face registration. Above characterization not considering the relationship between points, and therefore does not have to maintain a reasonable face structure. Active shape models (Active Shape Models, ASM) and active appearance models (Active Appearance Model, AAM) can express both the texture and shape (shape) are two characters.
Shape features of both by the point distribution model (Point Distribution Model, PDM) to express it. Figure 4 to 600 people face the statistical distribution of facial feature points in the image map, red dot denotes the mean of the points. Texture characteristic of each feature of the ASM is, respectively, generated by calculating the points around the neighborhood information corresponding to each feature of the response (the Response Map). Delineated in blue area in Figure 5 is used to calculate the response, and red dots indicate the actual facial feature point location. AAM uses the whole face to describe texture by facial feature point location is transformed to the standard shape is obtained regardless of the shape of face texture, and is independent of shape based on the principal component analysis method for modeling facial texture.
Depth research on human face image registration
Starting from 2006, deep neural network has been in computer vision, speech recognition, and natural language processing and other fields have achieved unprecedented success, also face registration brought a breezy spring. Scholars are no longer making a face descriptors of the building complex. Academic industry accepted method for deep face registration there are two classes: cascade Convolutional network face registration (Cascade CNN) depth face registration and multitasking.
As shown in Figure 6, Cascade CNN consists of three levels, each level contains multiple convolution network. First-level estimate is given an initial position, on the basis of two fine adjusted after feature point location. Multi-task distribution, Brigadier General registration and other attributes related to face training simultaneously. Property contains the head pose associated with facial features, facial expressions, such as smiling mouth is likely to be open, positive face feature points symmetrically distributed. Multitasking helps enhance the feature point detection accuracy. However the speed of convergence of the different tasks have different and difficult, difficult to train. Currently offers two solutions for adjust the training process of different tasks: early termination rule (task-wise early stopping criterion) and dynamic control mechanism.
Figure 6 Cascade CNN network model
3. face different scenarios face registration registration
Academics face registration with each passing day, industry product applications technical requirements have become more sophisticated, and face different scenarios put forward different requirements of registration.
Face recognition business the core problem is semantic alignment between facial image pixels, namely face feature point location. Error description of the feature location results in extraction of face feature severely deformed, leading to identify performance degradation. In order to better support for face recognition, we increase the scope of change the face frame, in order to reduce dependence on the face detection frame size. Face feature point we choose five, ensures a certain degree of face structure description, and reduces the effect of registration error on face recognition.
Figure 7 for face recognition
Beauty takes facial feature points reach ultra high precision positioning, such as eye makeup eyeliner lashes only positioning is accurate enough in order to achieve the natural fit beauty effect. In order to provide precision, we use the cascade model, rough face facial features, facial for fine positioning.
Figure 8 intelligent beauty
Facial self portrait effect using mobile video, the registration process is strict. Traditional face matching technology does not have the capacity to determine whether tracking success, in order to avoid lost phenomena in the process of tracking (tracking to the inhuman face), you must rely on time-consuming face detection, face registration we face judgment features have been added, reduce reliance on face detection. We use elongated depth of neural networks, and compression using SVD decomposition model and algorithm acceleration, model size control in the 1M, the processing time on mainstream mobile phones only 3ms. Model size and speed are the industry's highest standards.
1 face self-timer video special effects
Excellent face registration renewal
Excellent laboratory continuous follow up on technology trends, update versions. Face matching technology migration from traditional methods to the deep learning method, from the latest academic research results to the best choice, we went through several rounds of iterative update, did a lot of innovative and try it. Face registration version 1.0 was released in April 2013, rough face features, 4 months after version 2.0 of precision positioning was successfully published, and fun products. After version 3.0 accuracy greatly improved, while landing in beauty products. Version 4.0 using the deep learning method, precision has been further improved, with an average above the artificial level. In May this year we released the latest version 5.0 using depth multi-task learning method, speed and depth of network size has been greatly optimized mainstream mobile frame rates over 200, model 1M, and bring their own face judgment function. Mentioned in the introduction of face self-timer effect with the support of this version.
4. the follow-up research and development plan
Future we strive to enhance the user experience of applications fell, on the other hand is also actively exploring new scenarios. Currently face self-timer video registration and tracking defects still exist. To resolve this problem, improve the user experience depends on further ways of enhancing the human face registration of stability and accuracy. Other than applications referred to in this article has been, face matching technology can also be applied to intelligent access control system, financial core, live on the Internet industry, and many other fields. In new areas of application, studies how face matching technology to meet the new requirements is another issue that we will face.
Lei Feng network Note: this article by Lei feng's network, for reprint, please contact the original author and to indicate the source and author, no deletion of content.