摘要
C3D神经网络的提出极大地促进了视频动作识别技术的发展。但是,原始的C3D神经网络识别准确率不高,有较大的提升空间。针对准确率不高的问题,文章针对C3D神经网络的激活函数和梯度优化算法进行了优化,用更加平滑的Mish激活函数代替了原先的ReLU激活函数,用参数变化更稳定的Adam优化算法代替了传统的SGD算法。以此为基础提出了MA-C3D神经网络,我们将改进后的模型放到UCF101数据集上进行了测试,发现相比于一般的C3D神经网络,模型的准确率提高了7.4%。
C3D neural network has greatly promoted the development of video action recognition technology However,the origi⁃nal C3D neural network recognition accuracy is not high,there is a large space for improvement.To solve the problem of low accuracy,this paper optimized the activation function and gradient optimization algorithm of C3D neural network.The original ReLU activation function was replaced by a smoother Mish activation function,and the SGD algorithm was replaced by an Adam optimization algorithm with more stable parameter changes.MA-C3D neural network was proposed in this paper.We tested the improved model on UCF101 data set and found that compared with the general C3D neural network,the accuracy of the model increased by 7.4%.
作者
钱闻卓
Qian Wenzhuo(Hangzhou Dianzi University,Hangzhou 310018)
出处
《现代计算机》
2021年第35期70-74,94,共6页
Modern Computer
关键词
动作识别
识别准确率
三维卷积神经网络
action recognition
recognition accuracy
3D convolutional neural network
