Semantic segmentation network based on multi-scale cross attention feature fusion
Aiming at the problems of the DeepLabv3+semantic segmentation model only fuses single-scale low-level feature in decoder,and the fusion effect of high-level and low-level features is poor,which leads to low target segmentation precision,based on attention feature fusion (AFF )structure and DeepLabv3+network,a CAAF-DeepLabv3+segmentation network is proposed.Firstly,the network introduces multi-scale low layer features at different stages to optimize the spatial position information.Secondly,AFF is improved in a cross way to obtain a multi-scale cross-attention feature fusion(CAFF)structure,which can improve the information interaction between features,it also enhances salient features by learning the importance of high-level and low-level features on the channel,so as to overcome the problem of feature fusion with different semantics and scales and obtain fusedfeatures with high resolution and high semantic information.The results of training and testing on traffic road marking datasets show that,when the contour of the target is complex and more small size of targets are distributed,compared with UNet,PSPNet,DeepLabv3+,MobileNetv2-DeepLabv3+and AFF-DeepLabv3+networks,this network has the highest MIoU and MPA values,and the missed segmentation and false segmentation are obviously reduced.
NETL
NSTL
万方数据
