A Method for Multimodal Face Recognition by Facial Depth Prediction

Abstract

The invention discloses a method for multimodal face recognition through face depth prediction, the steps of which are: data extraction; face segmentation and scaling of Depth modal data; cascade network model training; face recognition network training ; model fusion; testing phase. The people's face depth image predicted by the network model in the present invention is not only similar to the real people's face depth image, but also the cascaded classification network increases the distance between the classes of the predicted depth image, so that the predicted people's face depth image has Richer identity information. The Depth mode of the face is predicted by the RGB mode of the face, which increases the diversity of the face mode data, and can realize multi-modal face recognition without changing the existing RGB camera hardware. Combining the RGB modality with the predicted Depth modality, the accuracy of multi-modal face recognition is higher than that of single-use RGB modality data.

Description

technical field [0001] The invention relates to a face recognition method, in particular to a multi-modal face recognition method through face depth prediction. Background technique [0002] RGB (red, green, blue three-channel color) single-mode face recognition technology has reached the bottleneck, but under uncontrollable conditions, such as changes in facial posture, especially changes in illumination, the RGB face Recognition accuracy still has a big impact. With the emergence and popularization of Kinect and RealSense cameras, people can obtain face data of more modes other than RGB, such as Depth (depth) and NIR (near infrared). Face data in Depth and NIR modes are not affected by light, and accurate face information can be obtained even in the dark. Therefore, combining the two modalities of RGB and Depth for multimodal face recognition can greatly improve the robustness to illumination. However, the currently widely used cameras are still ordinary RGB cameras. If...

AI Technical Summary

Problems solved by technology

This technical method has the following defects: it uses various modal information of the face, but these information are collected from the corresponding equipment, and the hardware equipment requirements are relatively high in the actual face recognition application; and in this technology, except RGB The face positioning of other modalities except the RGB mode is obtained from the face positioning mapping of the RGB mode, so that when the face positioning of the RGB modal data is inaccurate or the face positioning cannot be performed (it cannot be detected in a dark environment Face positioning in RGB mode), the face recognition effect of other modes is limited by the face recognition effect of RGB mode

Images