Location and detection of facial feature points

The first two weeks of research on facial expression recognition of the article, many methods are based on the extraction of facial feature points based on the completion, and then use the mesh model or deformation model to analyze and detect the feature points and classification methods to realize the recognition of facial expressions, It can be seen that the extraction and location of key points (feature points) in facial expression recognition occupies an important position, decided to draw 1-2 weeks to study some of the key points extracted from the face of the article, this week, the main research on the face feature point detection, positioning and face calibration algorithm:

The method of face feature point detection mainly looks at the ASM (Active Shape model) algorithm, is based on the point distribution model PDM algorithm, in recent years CVPR many articles have introduced, mainly also includes the training and the search two parts, the training sample only needs to include the human face area to be possible, There is no need to consider the image size problem, but it is necessary to manually record the K features of each training sample, the main point is to record the location of key features of the coordinate information, saved in a text file, you can consider the use of coding to achieve. Then is constructs the training set the characteristic vector, for example n the training sample, obtains the n characteristic vector, constructs the characteristic vector after the shape normalization, the goal is lies in the front manual calibration human face characteristic point to align, eliminates the different angle, the different distance, the different posture and so on external factors caused the non-shape interference, So that the model is more effective, generally this step is to use the Procrustes method is normalized, the point distribution model is also the appropriate translation, scaling and other transformations, without changing the distribution model to get the same point distribution model. The Procrustes method is also a method to solve the transformation matrix, obtains the last average face model, then takes the model to the PCA processing, selects the first several eigenvector. Then the local feature is built for each feature point, and the feature vectors are obtained by selecting M pixels on the feature points, and the local texture is obtained by the gray derivation of the vectors. The texture mean and variance are obtained, and the similarity measure between the new feature of a feature point and its well-trained local feature can be expressed by Markov distance. The second step is the ASM search, and the new position of each feature point is computed mainly by the affine transformation rotation scaling operation. The whole ASM algorithm is more clear in operation.

Also saw the ECCV2014 on the face detection and calibration of the article as well as the CVPR2014 above the use of local two-valued features of the human face registration algorithm. The main cascade detector in the previous article classification, many methods have been used in previous methods, including regression function, feature point regression algorithm and State-of-the-art method are directly adopted, but from the detection effect is very good, completely understand the algorithm, but also need to learn the previous basic algorithm, It is mainly the implementation and flow of basic algorithms related to machine learning. The method of the previous article is also introduced in the following article, the second article is of high application value, belonging to the current method of discriminant shape regression, mainly introduce the linear regression and regression tree (what is the regression tree do not know?) ), the overall idea first initializes the image to 1 shapes (the shape is a set of points, the goal is to move them to the corresponding eye nose mouth) and then based on this shape to calculate the surrounding pixels of each point, or two shape two points of the median pixel value (for the illumination of the robust, is usually the difference between two pixel pixels, this feature is features.

The difference between the current shape and the manually marked shape is then calculated delta_shape, then a function y = f (x) is trained, so delta_shape = f (Features). Finally, adding this delta_shape to the original shape is the final required face shape. This process is the face Alignmeng core process of the method. The so-called registration is the calculation of this increment. Training is the relationship between learning characteristics and this increment. The method of this paper is to cascade the process, which reduces the difficulty of each registration. Take the last result of the above into the first step loop 10 times. is the whole process. Is the whole idea:

The article does not describe how to extract pixels around a landmark. The rotation and dimensional transformations between the current shape and the average shape should be obtained first, after the transformed offset to the pixel point (detailed in one millisecond face Alignment with anensemble of Regression trees).

Reference documents:

[1] Joint Cascade face Detectionand ALIGNMENT.ECCV 2014.

[2] face Alignment at + FPS viaregressing Local Binary FEATURES.CVPR 2014.

[3] Cascade Pose regression.cvpr2010.

Location and detection of facial feature points