基于卷积神经网络的半身裙款式特征分类识别

Classification and Recognition of Bust Skirt Style and Common Features Based on Convolutional Neural Network

针对服装特征分类识别不够全面、较多分类特征导致效果较差的问题,提出一种带有Inception v2模组的快速区域卷积神经网络模型的女装半身裙多特征分类识别方法。建立一个包含8类款式、11种颜色、5种图案、4种长度,共计28种类别标签的女装半身裙样本库;以快速区域卷积神经网络(Faster r-cnn)结构为基础,引入一个Inception v2模组,对半身裙的款式及多种特征进行学习训练,通过全连接层将来自Faster r-cnn主干网络和Inception v2的分类信息进行特征融合并共享损失,以提高算法的准确率;将目标检测框与分类结果一起输出,在对半身裙图像精准定位的基础上实现了半身裙款式及常见特征的分类识别。结果表明:该方法的平均分类准确率为92.8%,可以有效地对女装半身裙款式、特征进行分类识别,并且可用于实际场景的服装图片中。

To solve the problem that the classification and recognition of garment features are not all-round enough and many classification features lead to poor effect,a multi-feature classification and recognition method for women’s bust skirt based on fast regional convolutional neural network model with Inception v2 module is proposed.A sample library of 28 kinds of women’s bust skirts is established,which includes 8 styles,11 colors,5 patterns and 4 lengths.Based on the structure of fast regional convolutional neural network(Faster R-CNN),an Inception v2 module is introduced to train the learning of the styles and multiple features of bust skirt.Through the fully connected layer,classification information from the faster R-CNN backbone network and Inception v2 has feature fusion and shares loss,to promote the accuracy of the algorithm.The target detection framework is output together with the classification results,which achieves the classification and recognition of bust skirt style and common features on the basis of accurate positioning of bust skirt images.The results show that the average classification accuracy of this method is 92.8%,which can effectively classify and recognize the styles and features of women’s bust skirts,and can be used for garment pictures in real scenarios.