基于增强特征表示的语义分割神经网络

Semantic Segmentation Neural Network Based on Feature Representation Augmentation

语义分割是计算机视觉领域中一个基础而重要的话题,针对语义分割中边界分割困难的问题,提出了一种利用类别整体特征以增强模型表征能力的语义分割神经网络结构。通过分析同类别内各像素特征应具有相似性、不同类别内的特征应具有可分性的特点,利用区域提议汇聚各类别区域内的特征,并使用关注特征的方法建立像素点与类别之间的联系,从而增强模型表征能力。通过在公开数据集上的实验分析比较,上述结构能有效提升像素点特征表示能力。

Semantic segmentation is a basic and important topic in the field of computer vision.Towards the diffi-culy of boundary segmentation in semantic segmentation,this paper proposed to use the features of categories to aug-ment the feature representation.Based on the truth that the feature of each pixel in the same category area should be similar,and the feature in different categories should be discriminative,we propose to aggregate features of each cate-gory through regional proposals and establish the relationship between pixels and features of categories using focal fea-ture method,so that the model representation capabilities can be enhanced.Through the experimental analysis and comparison of the public dataset,the structure can effectively improve the pixel feature representation ability.