Weakly supervised semantic segmentation algorithm based on double dimensional enhancement
Aiming at the high cost of fully supervised semantic segmentation,a weakly supervised semantic segmentation method using only low-cost image-level tags is proposed.Firstly,feature extraction is carried out on the input image through the backbone network,and a dual-dimensional enhancement module is inserted into the network to spread attention to areas that are not easy to focus on.Secondly,semantic similarity is learned by using the obtained feature map and self-attention module,and the similarity is further expanded to cover the area of the location map.Generate fine location map,and finally build the consistency of fine location map and initial location map to generate intensive and accurate location map in a self-supervised way.The average intersection ratio of 70.4%of the verification set and 69.2%of the test set was achieved on Pascal Voc 2012 using only image-level tags,which is better than other self-supervised methods.
NETL
NSTL
万方数据
