姿态检测网络在服装关键点检测中的应用

Application of posture detection network in key point detection of clothing

为了提供全面且精度较高的关键点用于确定服装的具体轮廓,提升服装配准和检索的精度,运用估计人体姿态的深度神经网络卷积姿态机(convolutional pose machine,CPM)建立模型,对数据图片进行色度等增强,利用高斯核函数建立图片真实标签,并且仿照特征金字塔改变前端网络结构,对模型进行训练。实验结果表明:卷积姿态机可以有效地应用于服装关键点检测;与其他模型相比,能够检测单个人全身的穿着及5种不同类别的服装,提升裤子或者衬衫等大种类的检测精度为1%。

In order to provide comprehensive and high-precision key points for determining the specific clothing contours and improving the accuracy of clothing registration and retrieval,this experiment introduced a convolutional pose machine(CPM),a typical method for estimating the posture of human body,to realize clothing detection.In the proposed model,the chroma of digital picture was firstly enhanced.Then,the Gaussian kernel function was used to establish the real label of the picture.Finally,the front end of network structure was changed according to the feature pyramid so that the model could be well trained.The results show that the convolutional pose machine can be effectively applied to the key point detection of clothing.Compared with other models,it can detect the whole body of a single person and five different types of clothing,improving the accuracy of pants or shirts by 1%.