Abstract: Nowadays, the number of clothing images on the Internet is growing rapidly, and the demand for intelligent classification of a large number of clothing images is increasing. R-FCN（Region-Based Fully Convolution Networks） is introduced into the garment image recognition. In order to solve the problems of long training time and low recognition rate of the deformed garment in the garment image classification, a novel improved framework HSR-FCN is proposed. Based on R-FCN, the new framework integrates the HyperNet network with the region proposal network, changes the learning mode of image features, so that the HSR-FCN can achieve a higher accuracy in shorter training time. Besides, the spatial transformation network is introduced into the model, and the transformation and alignment of input image and feature map are carried out, which strengthens the learning of multi-angle clothing. The experimental results show that the improved HSR-FCN network model effectively strengthens the learning situation of the deformed garment image, and in the case of shorter training time, the average accuracy of the original network model R-FCN is about three percentage points higher than that of the original network model, up to 96.69%.

Key words: garment images, deep learning, image classification, Region-Based Fully Convolutional Networks（R-FCN）, HyperNet, region proposal networks, spatial transformation networks

摘要： 目前网络上的服装图像数量增长迅猛，对于大量服装图像实现智能分类的需求日益增加。将基于区域的全卷积网络（Region-Based Fully Convolutional Networks，R-FCN）引入到服装图像识别中，针对服装图像分类中网络训练时间长、形变服装图像识别率低的问题，提出一种新颖的改进框架HSR-FCN。新框架将R-FCN中的区域建议网络和HyperNet网络相融合，改变图片特征学习方式，使得HSR-FCN可以在更短的训练时间内达到更高的准确率。在模型中引入了空间转换网络，对输入服装图像和特征图进行了空间变换及对齐，加强了对多角度服装和形变服装的特征学习。实验结果表明，改进后的HSR-FCN模型有效地加强了对形变服装图像的学习，且在训练时间更短的情况下，比原来的网络模型R-FCN平均准确率提高了大约3个百分点，达到96.69%。

关键词: 服装图像, 深度学习, 图像分类, 基于区域的全卷积网络（R-FCN）, HyperNet, 区域建议网络, 空间转换网络