Research On Weakly-Supervised Semantic Segmentation Method Of Remote Sensing Images Based On CAM

Remote sensing image semantic segmentation is the basic task of many advanced tasks such as target detection,element extraction,recognition and tracking.It aims to assign a category to each pixel in the image,which can be applied to multiple scenes such as precision agriculture,urban planning,earth observation,cartography and so on.At present,remote sensing image semantic segmentation methods based on deep learning rely on massive highquality datasets with manual pixel-level annotations for training.However,there are few data with pseudo-pixel annotations in remote sensing images,and manual labeling is timeconsuming and laborious.Compared with expensive pixel-level annotations,the data with classification annotation,bounding box and graffiti as labels is easier to obtain.Therefore,this topic uses image-level category labels as supervised information to study weaklysupervised semantic segmentation methods for remote sensing images.Currently,most weakly-supervised semantic segmentation methods are multi-stage methods,which based on Class Activation Map(CAM)to visualize pseudo-pixel annotations.The training steps of these methods are cumbersome.In addition,remote sensing images have the characteristics of different target scales and high background complexity,resulting in low quality of pseudo-pixel annotations,and the segmentation accuracy of the model still needs to be improved.To solve the above problems,this thesis proposes a weakly-supervised semantic segmentation method for remote sensing images based on CAM,which uses Res Net as the backbone network.The main research contents and work are as follows:(1)In response to improve the activation accuracy of small objects in remote sensing images and improve the pseudo-pixel annotations,this thesis designs a weakly-supervised semantic segmentation method for remote sensing images based on self-supervision learning.Through the scale equivariant regularization and equivariant difference detection mechanism,the network can learn by itself without adding additional supervision information.Finally,small targets in remote sensing images are activated accurately,and the edge of pseudo pixel annotation is further refined to narrow the gap between pseudo pixel annotation and real pixel annotation.(2)In order to reduce the tediousness of network training,this thesis further designs a single-stage CAM-based weakly-supervised semantic segmentation method for remote sensing images.First,this thesis proposes a CAM-aggregation method,which uses pseudopixel annotations for classification decisions to ensure the semantic integrity of CAM;secondly,this thesis proposes the loss of semantic refinement,and designs an end-to-end semantic segmentation framework using semantic refinement loss.This method is compared with the mainstream model on the remote sensing image data set,and the effectiveness of this model is verified through the result analysis.Compared with the traditional CAM algorithm,the quality of pseudo-pixel annotation has been significantly improved,and the overall MIo U score has increased from 48.76% to 51.68%,among which the small target "car" has improved greatly;experimental results show that this method has significant advantages over other methods in solving detailed segmentation in complex scenes.