【作者】 姚万业； 孙沛杰； 冯涛明； 王祝；

【Author】 Wanye Yao;Peijie Sun;Taoming Feng;Zhu Wang;Hebei Technology Innovation Center of Simulation & Optimized Control for Power Generation,North China Electric Power University;

【机构】 华北电力大学河北省发电过程仿真与优化控制技术创新中心；

【摘要】 卷积神经网络(Convolution Neural Network,CNN)凭借其高效图像分类处理能力近年来在人工智能领域运用广泛,但随着分类总数的增多,卷积核和神经网络层次相应递增使参数和计算量均增加,致使分类效率与准确率降低。本文提出一种新的卷积层结构,即空洞卷积金字塔(Dilation Convolution Pyramid,DCP)结构。通过3个并行的空洞卷积对特征层采样,将3个特征层加权融合作为下一层的输入。该层在增加少量参数的情况下拥有更大的感受野,在同等算力下多类别分类达到更高准确率。基于花数据集卷积网络分类结果表明,DCP结构的VGG-16和GoogLeNet网络在花卉品种分类的准确率相较于原网络分别提升1.862%和1.793%。更多还原

【Abstract】 Convolution Neural Network has been widely used in the fields of artificial intelligence in recent years due to its high efficiency in image classification processing.However,as the total number of classifications increased,the convolution kernel and the neural network hierarchy augmented accordingly.The geometric increase of the required parameters and the amount of computation decreases the efficiency and accuracy of classification.A new convolution structure,Dilation Convolution Pyramid(DCP),is presented in this paper.Three paralled void convolutions are used to sample the feature maps,and a weight fusion of the feature maps is used as the input for the next layers.This layer has a larger field of perception with a small number of parameters and a higher accuracy for multi-category classification under the same computing power.The accuracy of DCP-structured VGG-16 and GoogLenet networks in flower variety classification is 1.862% and 1.793% higher than the original network,respectively.更多还原