一种引入聚焦机制的静态图像人体分割方法

发布时间：2018-04-22 14:20

本文选题：深度卷积神经网络 + 残差网络　；参考：《电视技术》2017年Z2期

【摘要】：针对静态图像中人体分割不能够聚焦所要分割区域,造成分割效果不佳的问题。通过对残差网络进行改进,使之能够适应人体分割这一任务,并在改进的残差网络中引入一种聚焦机制进行静态图像人体分割。根据静态图像人体分割问题,设计了具有聚焦机制的损失函数、辅助分割损失函数以及分割损失函数。另外,为了得到数据集的辅助分割类标、分割类标以及注意力聚焦图,提出了数据预处理算法。在Pascal-Person-Part数据集和LIP数据集上进行训练和测试,并将测试结果与SegNet,FCN-8s,Deep Lab V2,Attention,LG-LSTM以及Attention+SSL方法进行比较。通过比较像素精度、平均像素精度和Io U(Intersection over Union)指标,表明所提方法能够提高静态图像中人体分割的效果,验证了所提方法的可靠性。
[Abstract]:In the static image, the human body segmentation can not focus on the region to be segmented, resulting in poor segmentation effect. The residual network is improved to adapt to the task of human body segmentation, and a focusing mechanism is introduced into the improved residual network for the segmentation of human body in static images. According to the human body segmentation problem of still image, the loss function with focusing mechanism, auxiliary segmentation loss function and segmentation loss function are designed. In addition, a data preprocessing algorithm is proposed in order to obtain the auxiliary segmentation class mark, the segmentation class label and the attention focus graph of the data set. Pascal-Person-Part data sets and LIP datasets were trained and tested, and the test results were compared with SegNetFCN-8s Deep Lab V2 tentiontionLG-LSTM and Attention SSL methods. By comparing pixel precision, average pixel precision and I o U(Intersection over union index, it is shown that the proposed method can improve the effect of human body segmentation in still image, and verify the reliability of the proposed method.
【作者单位】：郑州航空工业管理学院电子通信工程学院;河南广播电视大学信息工程系;
【基金】：国家自然科学基金项目(71371173) 国家航空科学基金重点项目(2014ZC13004) 河南省科技攻关项目(172102410067;144300510059;132102210474) 河南省教育厅重点研究项目(17A520065)
【分类号】：TP391.41

【相似文献】