摘要
为了缩短神经网络方法在人脸表情识别训练阶段所消耗的时间,提出了一种基于剪枝极限学习机的人脸表情识别方法.该方法随机为输入层与隐层单元间的连接分配权重,并以此求出相应的隐层单元与输出单元之间的连接系数,以达到一次性确定神经单元间系数的效果,避免了深度学习和神经网络因反复修改网络系数造成的高耗时;接着利用隐层单元与分类结果的相关性,删掉那些相关性较低的隐层单元,从而自动构建网络结构;通过多次迭代选取分类结果最好的网络参数作为最终的网络参数;最后构建多个极限学习机,用投票机制实现最终的分类.实验结果表明,该方法在特定人脸识别和非特定人脸识别都能得到较好的识别率,且大大降低了训练所用时间.
When a neural network was applied for facial expression recognition, it always took a lot of time in the training phase. In order to shorten the training time, a method based on the extreme learning machine for facial expression recognition was proposed in this paper. This method assigns weights between the hidden and input nodes randomly, and then finds the corresponding parameters of the connections between the hidden layer units and output units. In this way, it achieves the effect of determining the coefficients between nerve cells at one time, which avoids big amounts of time caused by repeated modifications by the deep learning and neural networks. Then based on the correlation between hidden layer units and classification results, it deletes the low relevance units in the hidden layer so as to build a network structure automatically. And through multiple iterations, network parameters with the best classification results are selected as the final network parameters. Finally more extreme learning machines are built using this new method and the final classification is achieved using the voting mechanism. The experimental results show that the method can achieve a relatively satisfactory recognition rate in specific and non-specific facial expression recognition and greatly reduces the training time.
出处
《五邑大学学报（自然科学版）》
CAS
2016年第2期35-40,共6页
Journal of Wuyi University(Natural Science Edition)
关键词
极限学习机
表情识别
模式识别
ELM
facial expression recognition
pattern recognition
