High-Resolution Remote Sensing Image Instance Segmentation Based On Deep Learning And Muli-Scale Fusion

Remote sensing image segmentation is an important method for extracting and identifying various ground objects.With the continuous improvement of remote sensing image resolution,first,the status of "same matter with different spectrum" and"same spectrum with foreign matter" in high-resolution remote sensing has been continuously strengthened,and pixel-based methods are prone to "salt and pepper effect";Secondly,the ground features in high-resolution remote sensing images have multi-scale features.Based on traditional object-oriented methods,such as the use of eCognition 's multi-scale segmentation method,it is prone to under-segmentation and over-segmentation.In addition,the demand for vector data in various industries has increased exponentially in recent years.Therefore,in the face of the ever-increasing demand for vectorization of high-resolution remote sensing images,how to meet the increasing needs and quickly extract and identify various features to complete the vectorized interpretation of high-resolution remote sensing images is the current main problem.Deep learning is also solving the above related problems.Among them,the instance segmentation based on deep learning can not only obtain the category information of the detected features than the semantic segmentation based on deep learning,but also distinguish different objects in the same category,which is more in line with the needs of geographic researchers.Interpretation results.Therefore,this paper aims at the problem of insufficient scale fusion in the segmentation of high-resolution remote sensing images.Taking a region in Jiangnan Town,Hangzhou City as an example,based on the Mask R-CNN instance segmentation model,a spatial data fusion scheme under different window scales is designed.The analysis of multi-scale fusion effects.The main research contents of this article are summarized as follows:(1)To achieve a vectorized interpretation scheme for high-resolution remote sensing images based on Mask R-CNN.First,through Mask R-CNN,the recognition and extraction of the main features of the high-resolution remote sensing image were realized;secondly,the automatic vectorization of the high-resolution remote sensing image was realized through the post-processing of the recognition data,and the more accurate vector with the boundary between the features Layers.In the experiment,the vector layers of the four main categories of roads,vegetation,residential areas and water bodies were obtained,and the automatic vectorization of high-resolution remote sensing images was realized.(2)To solve the topological ambiguity of the instance segmentation results of high-resolution remote sensing image.First,for the problem of edge information blurring caused by the mask R-CNN need to block input,an overlap strategy is used to supplement the context information;second,for the gap and overlap error generated by the Mask R-CNN model when predicting the mask,related problems are proposed.Topology repair method.In the experiment,a vector layer consistent with the high-resolution remote sensing image of the whole image was obtained,which verified the possibility that the example segmentation model can also be used in the application of the whole image segmentation.(3)To study the fusion scheme of multi-scale segmentation results,which obviously improves the completeness of road extraction.Aiming at how to comprehensively explore the effects of multi-scale,a multi-scale fusion scheme of different spatial windows is designed.In the experiment,taking the road as an example,it is proved that the segmentation result of the feature pyramid network of Mask R-CNN under multi-scale is better than the segmentation result under single-scale.