摘要
在采用深度学习进行图像分类时,为减少下采样导致的空间信息损失,往往采用膨胀卷积代替下采样,但尚未有文献研究膨胀卷积作用于不同网络层的性能差异。文中进行了大量图像分类实验,找到了适宜膨胀卷积作用的最佳网络层。但使用膨胀卷积会丢失近邻点的相关信息,导致网格现象,造成图像部分局部信息的丢失。为消除网格现象,又提出在前述最佳网络层采用多尺度膨胀卷积构建神经网络的方法。实验结果表明,所提出的构建网络方法在图像分类中取得了较好的效果。
In order to reduce the loss of spatial information caused by down sampling,dilated convolution is often used instead of down-sampling in image classification based on deep learning.However,there is no literature on the performance difference of dilated convolution on different network layers.In this paper,a large number of image classification experiments have been carried out,and the best network layer suitable for dilated convolution has been found.However,the use of dilated convolution will lose the information of neighboring points,resulting in grid phenomenon and the loss of partial information of the image.In order to eliminate the grid phenomenon,this paper also proposes a method of constructing neural network by using multi-scale dilated convolution in the optimal network layer mentioned above.The experimental results show that the proposed network construction method achieves good results in image classification.
作者
吴昊昊
王方石
WU Hao-hao;WANG Fang-shi(School of Software,Beijing Jiaotong University,Beijing 100044,China)
出处
《计算机科学》
CSCD
北大核心
2020年第S01期166-171,186,共7页
Computer Science
关键词
神经网络
图像分类
膨胀卷积
多尺度
Neural network
Image classification
Dilated convolution
Multi-scale
