基于SE-Inception v3与迁移学习的服装袖型识别与分类

Clothing sleeve recognition and classification based on SE-Inception v3 and transfer learning

为了提高服装袖型的识别与分类效率,提出一种融合SE(Squeeze-and Excitation)注意力机制和Inception v3主干网络的分类模型SE-Inception v3。针对图片背景等无关信息对识别的干扰问题,添加注意力机制,增强有用特征通道;引入了迁移学习思想,防止因袖型样本数据集较少而产生过拟合问题。将通道注意力和Inception模块多尺度卷积二者融合,有效地提升网络的特征提取和表达能力,该模型最终实现了以袖子为例的服装关键部位的识别与分类。通过对泡泡袖、灯笼袖、蝙蝠袖等8类服装袖型数据增强后共计3200个样本进行训练验证,平均准确率达到95.38%。与其它几类主流卷积神经网络模型进行对比实验,结果表明该模型具有较高的识别准确率,可为服装部位的图像分类识别提供有力支撑。

In order to improve the efficiency of garment sleeve recognition and classification,a classification model SE-Inception v3 was proposed,which integrates attention mechanism and Inception v3 backbone network.For the interference problem of irrelevant information such as the picture background,the SE(Squeeze and Excitation)attention mechanism was added to enhance the useful feature channel.The idea of transfer learning was introduced to avoid the over-fitting problem due to the small data set of sleeve samples.The fusion of channel attention and inception module multi-scale convolution effectively improves the feature extraction and expression capability of the network,and the model finally realizes the recognition and classification of the key parts of clothing with sleeves as an example.A total of 3200 samples were trained to verify the enhanced data of 8 types of clothing sleeves,such as bubble sleeve,lantern sleeve and bat sleeve,with an average accuracy of 95.38%.Compared with other mainstream convolutional neural network models,the results show that this model has high recognition accuracy and can provide strong support for image classification and recognition of clothing parts.