基于LBP特征和权重最优下的CNN人脸表情识别

CNN FACIAL EXPRESSION RECOGNITION BASED ON LBP FEATURE AND OPTIMAL WEIGHT

为了提高训练速度和浅层卷积神经网络下的人脸表情识别效果,提出一种基于LBP特征和权重最优下的卷积神经网络(CNN)人脸表情识别方法。通过对原始样本数据集进行预处理,并将处理好的数据集利用局部二值模式(LBP)获取局部纹理特征;将得到的LBP特征的样本数据集进行多次随机打乱,加入到自己创建的浅层CNN中,利用随机梯度下降算法(SGD)对每次打乱的样本数据集进行训练及验证,并保存最优权重下的CNN模型;用测试集对CNN模型在CK+数据集和FER2013数据集上进行10倍交叉验证并分别取得了97.2%和71.4%的识别率。实验结果表明,该方法不仅使网络训练迭代的速度变快,而且能有效增强人脸表情识别的效果,具有较强的鲁棒性。

In order to improve the training speed and the facial expression recognition effect under shallow convolutional neural network, this paper proposes a convolutional neural network(CNN) facial expression recognition method based on LBP features and optimal weight. The original sample dataset was preprocessed and the processed dataset was operated by local binary patterns(LBP) to obtain local texture features. The sample dataset with LBP features obtained was randomly scrambled several times and added to the shallow CNN created by ourselves. The stochastic gradient descent(SGD) algorithm was used to train and verify the sample dataset each time after it was scrambled, and the CNN model with the optimal weight was saved. The CNN model was 10-fold cross-validated on CK+ dataset and FER2013 dataset, and the recognition rates were 97.2% and 71.4% respectively. The experimental results show that this method not only make the speed of network training iteration faster, but also effectively enhance the facial expression recognition effect and has strong robustness.