Research On Satellite Image Segmentation Method Based On Deep Learning

With the rapid development of aerospace technology,the collection of satellite images has become easier and easier,and satellite images contain a wealth of information.How to quickly and efficiently extract the information people need from satellite images is an urgent problem to be solved.The use of manual processing of satellite images is not only time-consuming and labor-intensive,but also requires a high cost.Although traditional image segmentation methods can effectively segment the target,they are not universal,especially when the image is more complex,the segmentation effect Poor.With the rapid development of deep learning in recent years,more and more neural network models have been applied to tasks such as target recognition and image segmentation,such as satellite image segmentation,medical image segmentation,etc.Satellite images contain complex information and are difficult to process.The use of deep learning not only improves the processing results,but also saves a lot of costs.This article mainly applies the deep learning method to satellite image segmentation and makes improvements,the purpose is to effectively identify and segment the target in the satellite image.The main work of this paper is as follows:(1)Analyze common deep learning segmentation models,such as FCN,SegNet,YOLACT,Mask R-CNN.Two improvement schemes are proposed on the basis of Mask R-CNN.The first improvement scheme is to add a strip pooling module to the feature extraction backbone network of the original Mask R-CNN and modify the upsampling part of the prediction mask to Point Rend;improvement scheme two The feature extraction backbone network of the original Mask R-CNN is modified to Vov Net and the up-sampling part of the prediction mask is modified to Point Rend.In order to verify the effectiveness of the improved algorithm,the improved algorithm is trained and compared on the data set.The data set used in this article is divided into two parts.The data set used in experiment one comes from the RSOD-Dataset satellite image data set;The data set comes from the Mapping Challenge building data set.The experimental results show that although the improved Mask R-CNN takes 0.01 s more segmentation speed than the original Mask R-CNN,the segmentation accuracy has been significantly improved.(2)Based on the previous research in this article,design and develop a satellite image segmentation system.The system functions mainly include satellite image preprocessing,traditional satellite image segmentation,satellite image segmentation based on deep learning,model management and model training.