Research On Semantic Segmentation Methods For High Resolution Remote Sensing Images

With the rapid development of remote sensing technology,the acquisition of highresolution remote sensing images has become more and more convenient.High-resolution remote sensing images have higher detail expression ability and more spatial information,which can provide more accurate geographic information and more comprehensive geographic data.As a result,high-resolution remote sensing images are widely used in fields such as land use and land cover,urban planning,environmental monitoring,and agricultural production.Therefore,it is particularly important to perform intelligent interpretation for high-resolution remote sensing images,and remote sensing image segmentation technology is an important component of intelligent interpretation for remote sensing images.The segmentation result directly determines the reliability of intelligent interpretation for remote sensing images.Thanks to the rapid development of deep learning technology,especially the innovation of computer vision algorithms,remote sensing image processing has a new processing paradigm,which promotes the development of high-resolution remote sensing image segmentation tasks.This thesis combines an in-depth understanding of highresolution remote sensing images to study the application of deep neural network algorithms in semantic segmentation tasks for high-resolution remote sensing images.The main research content is as follows:(1)An attention-based alignment semantic segmentation network for high resolution remote sensing images.This method is dedicated to solving three important problems in high-resolution remote sensing image segmentation:foreground-background and foreground-class category imbalance problems,feature misalignment,and insufficient contextual information extraction.In order to obtain sufficient context information and stronger feature representation,this method designs a Contextual Augmentation Block(CAB)based on attention mechanism,which can apply different attention mechanisms for different features according to their semantic levels,resulting in enhanced features.Meanwhile,the method proposes a novel Feature Alignment Block(FAB)guided by anchor map,and the anchor map can generate more accurate position offsets.With the guidance of anchor map,FAB can better address feature displacement of objects in remote sensing images.The combination of these two blocks can make full use of multi-scale features and obtain stronger feature representations.In addition,based on annealing algorithm,annealing online hard example mining strategy(AOHEM)can be used to mine hard examples in the appropriate training stage,thus effectively alleviating the foreground-background and foreground-class category imbalance problems.The experimental results demonstrate that the proposed method achieves reliable performance on two challenging datasets and achieves a good balance between segmentation accuracy and algorithm complexity.(2)An Annealing Augmented and Aligned Semantic Segmentation Network for High Spatial Resolution Remote Sensing Images.Starting from the extreme example imbalance problem in high-resolution remote sensing images,this method re-designs the previous work to obtain a segmentation model that is more suitable for high-resolution remote sensing images.The entire model of this method is based on the assumption that the model can mine more valuable hard examples,because only the model trained on hard examples can better perform feature fusion and feature alignment.In the model optimization stage,the improved AOHEM training strategy divides the entire training process into two stages.In the first stage,the model focuses on improving generalization performance.After the model has good generalization performance and can accurately mine hard examples,high-confidence hard examples can be mined in the second stage,while avoiding the problem of model oscillation in the previous work.In terms of structural design,the improved CAB module can upsample features gradually through three different attention mechanisms,while capturing sufficient contextual information and performing information interaction among the features to be fused;Meanwhile,the improved FAB module conducts feature alignment through the guidance of salient features,which,compared to the anchor map in the previous work,not only contains spatial information but also channel information.With the guidance of salient features,FAB can generate more accurate position offsets for more precise feature alignment.It is worth mentioning that the AOHEM loss function,CAB module,and FAB module proposed in this thesis are universal for recognition tasks and can be easily inserted into other networks that use multi-scale features.The effectiveness of the proposed method is validated on three high-resolution remote sensing image datasets.The proposed method achieves state-of-the-art(SOTA)performance on all three datasets and surpasses the current best methods on two of them without relying on additional supervision and ensuring efficiency.