可变形有效感受野的人体图像语义分割算法

针对现有人体图像前景目标姿态变化、大小差异过大和边缘细节丢失等因素造成分割效果不佳的问题,提出了一种基于可变形有效感受野的人体前景分割算法.该算法将不同尺度的特征图进行融合,减少下采样过程中丢失的空间语义信息;同时结合可变有效感受野模块和边缘细化模块来捕获空间信息和语义信息,以适应算法对不同目标的有效感受野范围,并使有...

Body image semantic segmentation algorithm based on deformable effective receptive field

In order to solve the problems of poor segmentation effect caused by th e pose change, large size difference and edge detail loss of the existing body image foreground target,a body foreground segmentation algorithm based on deformable effective receptive field is proposed.The algorithm merges feature maps of different scales to reduce the spatial semantic information lost in the downsampling process;the variable effective receptive field module and t he edge refinement module are combined to capture spatial information and semantic information to i ncrease the effective receptive field range of the algorithm for different targets,and make the effective receptive field expand with the shape change of target attitude,size and so on;finally, focal loss is used to alleviate the imbalance between positive and negative samples.The experimental results show that, on the baidu people segmentation dataset,compared with other mainstream semanti c segmentation algorithms,the intersection ratio of the algorithm is as high as 88.45%,1.07% higher than the mainstream semantic segmentation algorithm deeplab V3+,3.71% higher than the c lassic algorithm U-net,and it has fast running speed,good stability,high timeliness and good robustness.