Abstract:

To address the problem that feature extracted by real time semantic segmentation model lacks contextual information, which causese inconsistent segmentation results between and within classes, a lightweight adaptive spatial attention model and channel attention model are proposed in this study. The adaptive channel attention module uses depthwise separable convolution to model the channel-level feature dependencies, adaptively adjusts the channel convolution kernel size, strengthens the contextual representation ability of high-level features, and enhances the intra-class consistency of segmentation results. The spatial attention module uses grouped convolution to obtain a larger flow area of feature information with a small amount of calculation, strengthens the contextual connection of features at the spatial level, enhances the spatial detail information of features, and enhances the inter-class distinguishability of segmentation results. Testing and analysis on the Cityscapes data set show that the lightweight contextual attention mechanism achieves 71.5% mIoU.

Key words: attention mechanism, semantic segmentation, multi-branch structure, information fusion, deep learning, grouped convolution, inter-class variation, intra-class consistency