Research On Remote Sensing Image Segmentation Algorithm Based On Deep Learning

At present,through the advancement of science and technology and the update of shooting methods,the quantity and quality of pre-satellite images have been greatly improved,and the information contained is also increasing.How to make full use of the images for research in the fields of environment,agriculture,and urban development.The main problem currently facing.In order to facilitate the research of predictive images,this subject has conducted in-depth research on instance segmentation in predictive images.First,the mainstream instance segmentation framework Pa Net is applied to the field of image segmentation,and a set of optimization methods are proposed for the shortcomings of Pa Net in processing remote sensing satellite images: the traditional watershed algorithm is improved,and the improved watershed algorithm is used to compare the previous Set the satellite image for preparation,reduce the difference between the front satellite image and the ordinary image,and improve the segmentation effect.At the same time,due to the extremely large spatial resolution in the expected image and obvious intra-class differences,Pa Net’s feature fusion of large-format images is insufficient,resulting in low detection accuracy and inaccurate segmentation.In response to the above problems,the Pa Net network structure was redesigned,so that the new network structure can better control the giant and giant satellite images,and a new feature adaptive branch is proposed.This branch supervises the feature fusion process of FPN.The fused features are classified into target sizes to further improve the pertinence of different size targets during feature fusion.Through experiments on different types of predicted images,the results show that our method is very competitive with the current state-of-the-art division method.