交互补充与验证的语义分割

Mutual Complementation and Validation for Semantic Segmentation

为了改善语义分割网络中特征的提取过程,明确模型的运行逻辑,辨别性特征被按照像素自身信息和像素间关联信息两个角度拆分,使用关系支路、信息支路两个支路网络专注于各自特征的提取。提出交互补充模块促进两个支路间特征的交互与补充,并利用交互验证模块对两个支路的预测进行额外验证。此外,在传统交叉熵损失的基础上提出对比交叉熵损失,使其更有利于对像素特征的约束与调整。在Dark-Zurich数据集和Cityscapes数据集上将该方法与不同的网络结合,进行大量实验验证了其有效性。

In order to improve the process of feature extraction in semantic segmentation network,and to clarify the model’s running logic,discriminative feature is decomposed into pixel information and pixel correlation.Information branch and relation branch are used to concentrate on the extraction of features respectively.Mutual complementation module is proposed to promote the interaction between branches,and the prediction of two branches could be verified by using mutual validation module.Besides,based on the conventional cross-entropy loss,contrasted cross-entropy loss is proposed to make it more conducive to the constraint and adjustment of pixel features.Extensive experiments are carried out on Dark-Zurich and Cityscapes datasets by combining this method with various networks to prove its effectiveness.