Research On Semantic Segmentation Methods For High-Resolution Urban Remote Sensing Images

Semantic segmentation of high-resolution urban remote sensing images allows us to obtain information of objects in the city,such as categories and locations.It is of great significance to city planning,disaster reduction,and even national defense.According to different shooting angles,high-resolution urban remote sensing images can be divided into ortho images and oblique images.In ortho high-resolution urban remote sensing images,there are usually huge scale variations not only among various categories,but also within a certain category.Therefore,the multi-scale features limit the performance of methods that process images in a single-scale manner.Moreover,the texture features of high-resolution images are complex,resulting in semantic category confusion.Different from the uniform spatial resolution in ortho images,the internal spatial resolution of oblique images changes significantly,which leads to the problem of blurred segmentation boundaries in semantic segmentation.And it is difficult to perform well on small object categories which are not involved in ortho images.To address these problems,the research is carried out from two aspects: ortho images and oblique images.1.Aiming at the problem of multi-scale features and complex textures in high-resolution ortho urban remote sensing images,a semantic segmentation method based on multi-scale features and attention mechanism is proposed.Based on the hierarchical features of Swin Transformer,by analyzing the effects of different feature fusion methods,a bilateral fusion feature pyramid structure is proposed.It fully integrates low-level features that retain detailed information and high-level features with rich semantic information.A parallel channel attention branch is also designed.The effect of attention mechanism on feature enhancement is discussed,and a channel attention module is proposed.Representations of different semantic categories are enhanced by assigning attention weights from the channel dimension.Experiments are carried out on two public remote sensing datasets of high-resolution urban scenes ISPRS Vaihingen and Potsdam.Experimental results have demonstrated the superiority of proposed network over state-of-the-art methods,reached 83.32% the mean intersection over union(m Io U)score on the Vaihingen dataset and 87.65% m Io U on the Potsdam dataset,respectively.This method realizes accurate and effective multi-category semantic segmentation of remote sensing images.2.Aiming at the problem of blurred boundary and low accuracy of small objects in oblique high-resolution urban remote sensing images,a semantic segmentation method based on boundary aggregation and small object mining is proposed.A boundary aggregation calibration module is designed to aggregate boundary information and recalibrate rough predictions,in order to locate object boundaries more precisely.To improve the performance on small object classes,a foreground-aware step is inserted,and a foreground perception loss function is introduced based on the dual-branch decoder structure.In the optimization stage,the small target mining strategy enables the model to automatically select hard samples in small targets,thereby optimizing the direction of network training.Combining multi-category segmentation main loss,boundary calibration auxiliary loss and foreground perception loss,an optimization strategy for multi-loss collaboration is formed.The experiments are conducted on the unmanned aerial vehicle oblique high resolution urban remote sensing UAVid dataset.The model achieved a m Io U score of 68.40%,which exceeds the existing methods.This method realizes high precision semantic segmentation of oblique remote sensing images.