基于DenseNet的图像识别方法研究

Research on Image Recognition Method Based on DenseNet

在深度学习中,随着神经网络层数的加深,训练网络变得越来越困难,现有的浅层的网络无法明显提升网络的识别效果。针对在原有网络的情况下提升网络识别效果,减少参数的问题,本文提出一种改进的DenseNet网络模型,借鉴Inception V3的思想,利用非对称卷积将DenseNet网络中Dense Block层所包含的3×3卷积分解成3×1和1×3的两个卷积,简化网络结构模型。之后再对改进前与改进后的网络模型在数据集上进行训练,实验结果表明,与经典的DenseNet网络模型相比,改进后的网络模型可以提高图像识别的准确性,降低参数,有效地缩短运行时间。

With the deepening of the layers of neural network in deep learning,the training network becomes more and more difficult,and the existing shallower network can not significantly improve the recognition effect of the network.In order to improve the network recognition effect and reduce the parameters in the case of the original network,an improved DenseNet network model was proposed.Based on the idea of Inception V 3,the 3×3 convolution contained in Dense Block layer in DenseNet network was decomposed into 3×1 and 1×3 convolutions by using asymmetric convolution,which simplifies the network structure model.The network was trained before and after the improvement on the dataset.The experimental results show that compared with the classical enseNet network model,the improved network model can improve the accuracy of image recognition,reduce the parameters and shorten the running time effectively.