Non-local network provides a pioneering approach for capturing long-range dependency by aggregating query-specific global context into each query location; however, non-local network applies the identical weight to each channel of feature maps and ignores the differences from the different channels of features. We design a novel tensor attention module (TAM), which integrates the context information along spatial dimension and channel dimension by introducing a bias learnable parameters tensor, so that the feature at each location of each channel can aggregate the features from all other locations. Motivated by SE-Net, we propose a novel second-order covariance attention module (SCAM) to enhance the feature correlation between different channel maps through the second-order statistics and the local cross-channel interaction strategy. We take the encoder–decoder segmentation network DeepLabv3+ as baseline, and in the encoder develop the attention modules TAM and SCAM for semantic segmentation (TCNet). Experimental results on PASCAL VOC 2012 and Cityscapes datasets show that our proposed network has better performance than the other state-of-the-art segmentation networks.

非本地网络通过将特定于查询的全局上下文聚合到每个查询位置，提供了一种捕获远程依赖性的开创性方法；然而，非局部网络对特征图的每个通道应用相同的权重，并忽略不同通道特征的差异。我们设计了一种新颖的张量注意力模块（TAM），通过引入偏置可学习参数张量，沿空间维度和通道维度整合上下文信息，以便每个通道每个位置的特征可以聚合所有其他位置的特征。在SE-Net的推动下，我们提出了一种新颖的二阶协方差注意模块（SCAM），通过二阶统计和局部跨通道交互策略来增强不同通道图之间的特征相关性。我们以编码器-解码器分割网络 DeepLabv3+ 作为基线，并在编码器中开发用于语义分割（TCNet）的注意力模块 TAM 和 SCAM。PASCAL VOC 2012 和 Cityscapes 数据集上的实验结果表明，我们提出的网络比其他最先进的分割网络具有更好的性能。