A Real-time Semantic Segmentation based on Self- attention and Contextual Information

Semantic segmentation methods use the backbone which does not have enough receptive field to deal with the image problem or has no contextual information to solve the blurred edge of images. Focusing on the problem and reducing parameters and computations, we improve a lightweight semantic segmentation method. In our network model, we use an improved D-block as the backbone. It utilizes a pair of parallel convolution layers with varying dilation rates to expand the receptive field, all the while incorporating an MCE block to maintain local details. The introduction of the Channel Context Module (CCM) aims to capture the enduring semantic relationships between channels and the mapping of semantic features. Subsequently, the Spatial Context Module (SCM) is employed to uncover the spatial context connections among pixels, by investigating the correlations between pixels and classes. Finally, the final result is obtained by multi-scale feature fusion. Our model gets good results on two datasets and achieves 76.1 mIoU on Cityscapes and 82.9 mIoU on CamVid without ImageNet pre-training.