Research On Remote Sensing Image Segmentation Based On Differential Pyramid And Edge Loss Enhancement

With the development of remote sensing technology,the number of remote sensing images increases rapidly,followed by a significantly raise in their spatial resolution.Parsing the effective information from these huge amounts of complicated data has become an important issue in this field.As an essential technique in remote sensing image parsing problem,semantic segmentation aims to determine the category as well as the precise location of object by performing the pixel-wise semantic labeling.As a result,such technique addresses wide application scenarios such as agricultural detection,navigation,urban planning,and related fields in military.Traditional approaches usually adopt hand-crafted features cooperated with individual classifiers to solve the remote sensing image segmentation problem.Unfortunately,such methods suffer from several drawbacks as weak representation ability of visual information,too many pre-defined parameters and low computational efficiency.As a branch in deep learning,convolutional-neural-network-based approaches have gained a good reputation in many visual applications such as image classification,image segmentation and object detection for their excellent representation learning ability.It is also gradually recognized as an effective solution in remote sensing image segmentation.However,due to the risk of confusing deep features between object and the background,the existing methods may involve several general issues as imperfect extraction of object region or boundary blurring.To address these problems,a comprehensive study is carried in the following aspects:1.We construct two remote sensing image datasets for planting greenhouses and photovoltaic panel components,respectively.While the images in the former one were captured in northern China and the images in the latter one were collected from Google Earth software.We performed regular cropping to ensure a fixed resolution of samples in both datasets as well as the manual annotation in a pixel-level manner.Finally all samples were assigned to the training and the validation sets following a certain proportion.2.Aiming at the problems of low segmentation accuracy and target boundary blurring issues,we proposed an edge loss enhanced semantic segmentation network that comprehensively utilizes the boundary information and hierarchical deep features.By introducing an edge loss enhancement structure to the dense feature extraction network structure,the regional semantics and edge information of the target are jointly learned.In addition,we also performed a comparative experiment on the remote sensing image datasets and the result illustrates that our approach improves the segmentation performance in edge locating perspective.3.Inspired by the observation of object characteristic in remote sensing images,we propose a novel salient target detection model based on the deep center-surround contrast pyramid.The dilated convolution is employed to construct the center and the surround pyramids which are used to capture the contrast information across different scales.As the proposed module is independent of the feature extraction layers,it can be seamlessly integrated into general net structures and ensure end-to-end training and inference.Extensive experiments on four popular benchmarks have proved the effectiveness of the method.The additional experiment on remote sensing image datasets also shows that the proposed method is superior to the competing approaches.In summary,this research for remote sensing image segmentation aims to explore the effective visual representation method for distinguishing the object and the background in deep feature level.To this end,approaches such as emphasizing the edge information and the local prominence are carefully studied.The extensive qualitative and quantitative experimental results have proved the effectiveness of the proposed algorithms.Therefore,we expect that our algorithms might have potential application prospects in the related area of remote sensing image segmentation.