基于深度学习的人脸图像扭正算法

An face imagetwist algorithmbased on deep learning

人脸识别作为目前最方便的生物特征识别技术,被应用到了很多重要的领域。但是,由于光照以及姿态等因素的影响,使得人脸识别的精度降低,造成人脸识别技术在实际应用中的局限性。针对姿态以及光照因素对人脸的影响,提出一种基于深度学习的人脸扭正算法。该算法将对齐后的人脸图像首先用深度卷积网络自动地提取人脸特征,然后根据提取到的特征得到非正面人脸与正面人脸的映射关系,最后将非正面姿态的人脸扭成正面姿态且处于中性光照下的人脸图像,算法引入了欧式距离与余弦距离两个损失函数来对网络进行优化,进一提高了网络的精度。实验结果表明,该算方法可以有效地实现正面人脸的重构,减少姿态与光照对人脸特征的影响,使人脸识别精度提得到提高。

As the most convenient, immediate and confidential biological recognition technology, face recognition has been ap-plied to more and more importaot fields. In consideration o f the influence of illumination, poseand other factors, the recognition accuracy often tends to be very low and may cause inaccurate recognition results. The paper proposes a face image twist algor-ithm based on deep leaming. Firstly, the deep convolution network designed was used to extract face feature automatically. Sec-ondly, the feature extracted from an image under any illumination and pose was used to obtain the mapping between non-frontal face and the frontal face. Thirdly, the face image in the canonical view and neutral illumination can be got. The Euclidean dis-tance and the cosin distance were introduced to optimize the network simultaneously, which improves the accuracy o f the net-work. The experimental results show that the algorithm can be used to reconstruct canonical face, and reduce the influence of the pose and illumination on the facial features, so the accuracy of face recognition can be improved.