基于残差网络的非限定条件人脸识别研究

Research on Unconditioned Face Recognition Based on Residual Network

针对非限制人脸识别中人脸图像的尺寸和角度影响识别精度的问题,本文根据渐进校准的思想,设计出一种以具有渐进校准功能的卷积神经网络为分析算法的人脸识别方法。首先在非限制环境下对人脸图像进行几何归一化处理,并且利用主成分分析法进行降维;然后基于仿射变换和局部人脸分割理论,提出基于细节变换与特征融合的方法对人脸进行矫正;最后利用残差卷积神经网络构建人脸识别模型,在LFW数据集上对模型参数进行训练,并对训练后的模型进行仿真和检验。实测表明,通过矫正得到的正面人脸图像虽然存在轻微的扭曲现象,但其提取的特征信息能够有效提高非限制条件下多姿态人脸的识别准确率。

Aiming at the problem that the size and angle of face image in non-restricted face recognition affect the recognition accuracy, this paper designs a face recognition method based on progressive calibration to construct a convolutional neural network with progressive calibration function. Firstly, the face image is geometrically normalized in an unrestricted environment, and the principal component analysis method is used to reduce the dimension. Then based on the affine transformation and the local face segmentation theory, a method based on detail transformation and feature fusion is proposed. The face is corrected. Finally, the residual recognition convolutional neural network is used to construct the face recognition model. The model parameters are trained on the LFW dataset, and the trained model is simulated and tested. The actual measurement shows that although the positive face image obtained by the correction has slight distortion, the extracted feature information can effectively improve the recognition accuracy of multi-pose face under unconstrained conditions.