基于多样化结构的轻量型卷积神经网络设计

Design of lightweight convolution neural network based on diversified structure

传统的卷积神经网络的网络结构单一或模块单一,网络提取特征时都是以同一种模式持续进行,导致分类精度不够理想,且模型参数和计算量较大。针对这一问题,提出一种使用多样化模块的轻量型卷积神经网络结构Diverse-net,使用Reduce Module和Slice Module两种卷积模块,通过交替叠加使用两种不同的卷积模块来代替传统网络的单卷积核结构,增加网络的深度和宽度,并在不同模块之间加入残差学习。使用文中搭建的卷积神经网络Diverse-net,与其他网络模型在GTSRB和101_food数据集上进行测试来对比网络性能与模型大小,Diverse-net网络模型大小减少至20.8 MB,在数据集GTSRB上识别率可达到98.72%;在数据集101_food上识别率可达到68.09%。实验结果表明,所设计的卷积神经网络Diversenet在图像分类方面性能更优,且网络的模型较小。

The network structure or module of the traditional convolution neural network is single,and the features extraction by the network is carried out in a same mode continuously,resulting in the undesirable classification accuracy and a large amount of model parameters and computations. Therefore,a lightweight convolution neural network structure Diverse-net with diversified modules is proposed,in which the two kinds of convolution modules of Reduce Module and Slice Module are used to replace the single convolution kernel structure of the traditional network by means of the alternating superposition,which increase the depth and width of the network,and add the residual learning between different modules. The convolution neural network(Diverse-net) constructed in this paper and other network models were tested on GTSRB and 101_food datasets to compare their network performances and model sizes. The Diverse-net network model is reduced to 20.8 M,and the recognition rate can reach 98.72% on GTSRB dataset,and 68.09% on 101_food dataset. The experimental results show that the designed convolutional neural network Diverse-net has better performance in image classification and smaller network model.