多尺度SE-Xception服装图像分类

Clothing image classification based on multi-scale SE-Xception

应用当前较新颖且分类性能靠前的卷积神经网络Xception作为基础网络结构,尝试采用多尺度的深度可分离卷积来提升模型特征信息的丰富度,在模型中嵌入SE-Net模块增强有用特征通道,减弱无用特征通道.实验结果表明:提出的多尺度SE-Xception模型在2种噪声程度不同的服装数据集中均取得不错表现;ACS数据集的平均分类准确率为78.34%,分别高于VGG-16、ResNet-50和Xception模型8.52%、4.81%、3.69%;验证了多尺度SEXception模型具有更好的特征提取能力,能够提取到更多的服装信息,从而提高服装图像分类效果,一定程度上解决了特征尺度单一、信息丰富度低的问题.

The current newer and better convolutional neural network Xception was used as the foundation network structures.Multi-scale depth separable convolution was employed to improve the richness of model feature information.The SE-Net model was embedded in the model to enhance the useful feature channels and weaken the useless feature channels.The experimental results show that the multi-scale SE-Xception model achieved good performance in two different noise clothing datasets.The average classification accuracy of ACS dataset was 78.34%,which was higher than VGG-16,ResNet-50 and Xception models by 8.52%,4.81%and 3.69%,respectively.Therefore,it’s verified that the multi-scale SE-Xception model has better ability to extract features,can obtain more clothing information,and thus improve the clothing image classification effect.To some extent,a multiscale SE-Xception network is conducive to solve the single feature scale and low information richness problems.