基于MBP算法和深度学习的人脸识别

为了解决在深度学习提取人脸图像特征时,易忽略其局部结构特征和缺乏对其旋转不变性学习的问题,提出了一种基于单演局部二值模式(MBP)与深度学习相结合的高效率人脸识别方法。首先,用多尺度单演滤波器对图像进行滤波,得到幅值和方向信息;其次,用LBP算法和象限比特的方法进行编码,分块计算组合其直方图特征;然后,将提取的单演特征作为深度信念网络(DBN)的输入,逐层训练优化网络参数,得到优异的网络模型;最后,将训练好的DBN网络在ORL人脸数据库上进行人脸识别实验,进行识别率计算,其识别率为98.75%。所提出的方法使用无监督的贪婪算法,隐藏层设定为2层,使用反向传播算法优化网络。相较于已知的人脸识别方法,MBP+DBN算法对光照、表情和部分遮挡变化具有较好的鲁棒性,在人脸识别中识别率较高,具有一定的优势,为图像特征提供了一种新的识别方法。

Face recognition based on monogenic binary patterns and deep learning

In order to solve the problem of ignoring its local structural features and lacking its rotation invariance learning when extracting face image features from deep learning, an efficient face recognition method based on Monogenic Binary Pattern (MBP) and deep learning is proposed. First, the image is filtered by using log-Gabor filter to obtain amplitude and monogenic direction information. Next, the LBP and the quadrant bit method are used for encoding, and the histogram feature is combined by block calculation. Then, the extracted monogenic features are used as the input of Deep Belief Network (DBN), and the network parameters are trained and optimized layer by layer to obtain excellent network model. Finally, the trained DBN network performs face recognition experiments on the ORL face database, and the recognition rate is 98.75%. The proposed method uses unsupervised greedy algorithm, the hidden layer is set to 2 layers, and the back propagation algorithm is used to optimize the network. Compared with known face recognition methods, MBP+DBN algorithm has better robustness to illumination, expression and partial occlusion changes, and has higher recognition rate in face recognition. It has certain advantages and provides a new recognition method for image features.