A novel fusion of dense-sparse representation and large kernel attention mechanism for semantic segmentation network of street view images

In this paper, we propose a street view image segmentation network that combines dense-sparse representation with large kernel attention mechanism. Aiming at the problem of rapid accumulation of sensory fields in semantic branches, the dense convolution combines strip convolution and expansion convolution of adaptive sensory fields in a ‘dense-sparse-dense‘ way, and the unilateral kernel attention mechanism is integrated in the new branch to enhance the screening of background noise; moreover, in the deep feature extraction module at the intersection of the two branches, the sparse representation of the kernel expansion convolution is introduced. The sparse representation of the large kernel expansion convolution is introduced to integrate the global and local information. Experiments show that the proposed segmentation network effectively improves the global modelling capability of the network and reduces the background noise of the large kernel convolution, obtaining 78.8% and 81.8% mIoU on Cityscapes and CamVid datasets.