改进SSD的服装图像识别方法

Clothing image recognition method based on improved SSD

针对因形变引起的服装图像识别准确率低的问题,提出一种基于改进SSD的服装图像识别方法。以SSD为基础模型,首先采用空间变换网络对服装图像特征进行数据增强,通过仿射变换得到适合于服装图像的特征图,增强空间不变性,减少数据变形的偏差;然后采用可变形卷积改进标准卷积,采用偏移方法使特征抽取过程聚焦于有效信息区域,实现更加准确的特征提取;最后通过几何信息融合网络获得更多的几何细粒度特征信息,提高网络对几何信息的表达能力。结果表明:与原SSD模型相比,所提方法能显著提升对形变服装图像的识别准确度,其平均预测精度m AP值提升3.63%。

In view of the problems of low accuracy in clothing image recognition due to deformation,a clothing image recognition method based on improved SSD was proposed.Based on the SSD model,firstly,a spatial transformation network was adopted for the data enhancement of clothing image features,and a feature map suitable for clothing images was obtained by affine transformation,the spatial invariance was enhanced,and the deviation of data deformation was reduced.Then,deformable convolution was employed to improve the standard convolution,the feature extraction process was focused on the effective information regions by using the offset method,and more accurate feature extraction was achieved.Finally,more geometric fine-grained feature information was obtained through a geometric information fusion network,and the expression ability of the network for geometric information was improved.The experimental results show that the proposed method can significantly improve the recognition accuracy of the deformed clothing images compared with the original SSD model,and the average prediction accuracy m AP value can be increased by 3.63%.