Method for automatically detecting three-dimensional facial mark

Abstract

The embodiment of the invention provides a method for automatically detecting a three-dimensional facial mark, and relates to the technical field of machine learning and pattern recognition. The detection method comprises: roughly detecting a nose using a three-dimensional local shape descriptor; accurately positioning the tip of the nose by utilizing the characteristics such as the symmetry of the human face; determining areas where eyes and a mouth are located according to the distribution of the human facial features; using a convolutional neural network to minimize combination loss, and providing candidate objects of canthus and mouth corners; and maximizing the similarity between the candidate points and the real marks according to the features of the candidate points and the adjacent points thereof, and iteratively updating the candidate objects to realize accurate detection of canthus and mouth corners. According to the method, under the condition that only three-dimensional face data are used, high-precision automatic face mark detection is achieved, and the method is stable for facial expressions and postures and higher in practicability.

Description

technical field [0001] The invention relates to the field of human face pattern recognition, in particular to a method for automatically detecting three-dimensional facial landmarks. Background technique [0002] Accurate facial landmark detection is very important for many scientific researches and applications, such as: face recognition, face animation, expression recognition, object tracking, etc. Two-dimensional face images provide researchers with rich texture information, but are very sensitive to background, lighting, posture, etc. Whereas 3D face-based landmark detection methods use the 3D coordinates of the face and are robust to both uneven illumination and changing poses. Most of the current 3D facial landmark detection methods require texture as an additional input, but in the actual acquisition scene, the texture is not necessarily completely consistent with its 3D facial data, and even some facial 3D reconstruction methods cannot provide texture data at all. ...

AI Technical Summary

Problems solved by technology

Two-dimensional face images provide researchers with rich texture information, but are very sensitive to background, lighting, posture, etc.

In addition, in real scenes, the rigid deformation (i.e., pose change) and non-rigid deformation (i.e., expression change) of the detected object also makes the detection more challenging.

Images