一种引入聚焦机制的静态图像人体分割方法

摘要

In view of the human parsing in static image,previous methods cannot focus on the regions that segmented.This will make the performance of these methods become bad.A mechanism is introduced into the proposed method which can focus regions segmented.ResNet is used in the proposed method and adapted to the task of human parsing.According to the problem of human parsing in static images,the loss function,auxiliary loss function and the loss function of the focus mechanism are designed in this paper.In addition,in order to get the auxiliary segmentation labels of the data sets,the segmentation labels and the attention focus map,a data preprocessing algorithm is proposed.Experiments are conducted on Pascal-Person-Part dataset and LIP dataset.Compared with SegNet,FCN-8s,DeepLabV2,Attention,LG-LSTM and Attention + SSL the experiment results show that the proposed method can achieve better human parsing performance.The results of pixel accuracy,mean pixel accuracy and IoU metric indicate the effectiveness of our method and our method can improve the human parsing results.%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,DeepLabV2,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IoU(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.