Lightweight Real-Time Image Semantic Segmentation Network Based on Multi-Resolution Hybrid Attention Mechanism

Table 5

To verify the impact of each module on the final accuracy of our network, we validate the mIoU results on the Cityscapes set. “” represented the network use TensorRT to speedup.

Model	Resolution	GFLOPs	Parameters	FPS	mIoU
ERFNet [16]	5121024	—	2.1 M	41.7	68.0
DABNet [7]	5121024	—	0.76 M	104.2	70.1
ASFNet [8]	5121024	15.35	5.42 M	185	70.9
FRFNet-slim [32]	5121024	11.38	—	206.3	65
FRFNet [32]	5121024	16.01	—	132.7	69.5
STDC1-Seg50 [33]	5121024	0.81	8.4 M	250.4	71.9
STDC2-Seg50 [33]	5121024	1.44	12.5 M	188.6	73.4
BiseNetv2 [18]	5121024	21.1	—	156	72.6
MHANet(ours)	5121024	14.25	5.42 M	203	71.87