基于改进的深度卷积神经网络的人脸识别

摘要

Considering that many traditional deep convolutional neural network methods suffer from ineffective performance in face recognition under non constraint condition,and inspired by multi-scale and multi-feature fusion in image recognition,a face recognition method based on multi-scale feature convolutional neural network was proposed.The method was used to extract the features of the input image using two sub convolutional networks with different sizes.The object classification of the model was used to extract the feature of single layer from different feature extraction results,and the method was tested on the face databases of YouTube and AR.Results show that compared with the original structure method,the recognition rate of the improved structure improves obviously,which provides a solution to the limited number of feature sources of the CNN model.%针对传统结构的深度卷积神经网络对非约束条件下的人脸识别性能较差等问题,受到人对图像识别时多尺度多特征融合的启发,提出一种基于多卷积多尺度特征的人脸识别方法.对输入的图片使用两个相对独立不同粒度的子卷积网络进行特征提取;对于模型的对象分类,采用来自不同特征提取单层的特征提取结果,在YouTube和AR等两个人脸库上对所提方法进行实验.实验结果表明,与传统结构系统相比,改进结构后,系统的识别率有了明显提高,为传统的CNN模型层数过深时,特征来源受限提供了一种解决思路.