结合编解码网络的人体解析方法

Human Parsing Method Combined with Encoder-decoder Network

人体解析是语义分割的一个子任务,只对图片中的人物进行分割而忽略背景信息.人体解析任务由于其复杂性,导致现有网络分割不够精确.本文针对该情况提出了一种编解码网络.在编码器中,对特征提取网络的下采样倍数进行调整以得到合适分辨率的特征图.在解码器中,通过金字塔池化网络来提取上下文信息,并采取空间加通道的双注意力模块来修正特征图.本文的网络与经典的编解码网络在公开的人体解析数据集(LIP)上进行了对比,较Unet提升了10.70%MIOU,较Deeplabv3+提高了1.93%MIOU.实验结果表明,本文网络的特征提取能力以及解析能力更加适合人体解析任务.

Human parsing is a sub task of semantic segmentation,w hich only segments the characters in the image and ignores the background information.Due to the complexity of human parsing,the existing netw ork segmentation is not accurate enough.This paper proposes an encoder-decoder netw ork for this situation.In the encoder,the dow n sampling multiple of the feature extraction netw ork is adjusted to get the appropriate resolution of the feature map.In the decoder,the pyramid pooling netw ork is used to extract the context information,and the dual attention modular of space plus channel is used to modify the feature map.The netw ork in this paper is compared w ith the classic encoder-decoder netw ork on the open human parsing dataset(LIP),w hich is 10.70%more M iou than Unet and 1.93%more M iou than deeplabv3+.The experimental results show that the feature extraction ability and analytical ability of this netw ork are more suitable for human parsing.