摘要
针对LeNet-5在图像识别上的局限性,提出了一种改进的图像识别卷积神经网络结构,使其能具有更高准确率的同时具有更快的处理速度。首先使用BN方法对输入数据进行批规范化,再对卷积核进行分拆,并构建了更深的网络,去除全连接层,改用平均池化代替。最后进行验证,实现了对LeNet网络的改进。
Aiming at the limitations of LeNet-5 in image recognition, an improved image recognition convolutional neural network structure is proposed to make it have higher accuracy and faster processing speed. The BN method was used to batch normalize the input data, and then the convolution kernel was split, and a deeper network was built. The full connection layer was removed and the average pooling was used instead. Finally, verification was implemented to improve the LeNet network.
作者
吴阳阳
彭广德
吴相飞
Wu Yangyang;Peng Guangde;Wu Xiangfei(School of Electro-mechanical Engineering, Guangdong University of Technology, Guangzhou Guangdong 510006, China)
出处
《信息与电脑》
2018年第7期127-130,共4页
Information & Computer
关键词
卷积神经网络
BN方法
池化
convolutional neural network
BN method
pooling