Dual attention-guided cross-layer optimised semantic segmentation algorithm for street scenes

Aiming at the problems of slow speed of semantic segmentation of street scenes, unsmooth segmentation edges, and difficulty in accurate segmentation of small targets, this paper proposes a dual-attention-guided cross-layer optimised semantic segmentation algorithm for street scenes. Firstly, the backbone network is replaced and modified to improve the segmentation speed; secondly, a cross-layer feature fusion decoding network based on spatial alignment is constructed to achieve the fusion of semantic information and detail information, and to enhance the expressive ability of targets at different scales; lastly, achannel and spatial attention mechanism is proposed to model the global channel correlation and long-distance location correlation, and to enhance the network's ability of learning key features. The experimental results on the street scene dataset Cityscapes show that the proposed feature extraction coding network, cross-layer feature fusion decoding network, and attention mechanism module are effective; the experimental results show that the training speed of the model is improved by 2-3 times with the image of 2048×1024 size as the input, and at the same time, the test accuracy of the mIoU can reach 73. 8%, and the test speed is improved by 1.340s/sheet. 1.340s/picture to 0.689s/picture, and the problem of incomplete segmentation and mis-segmentation of ground traffic signs is effectively improved.