Research And Design Of Large Scale Satellite Image Semantic Segmentation System

The semantic segmentation of satellite image refers to labeling useful information in the satellite image.Satellite image contains plentiful information because of its largescale,it is necessary to extract useful information from large-scale satellite images by image segmentation algorithm.However,existing image segmentation algorithms only support small-scale image segmentation,and cannot directly segment the large-scale image.In order to solve the problem of large-scale image segmentation,this thesis researches and designs a large-scale satellite image semantic segmentation system,which explores from the following three aspects.Firstly,the large-scale image is cut into many small images so that they meet the image segmentation algorithm's input requirements.Because the cropped small image may lose part of the feature information in the image boundary,which leads to the false recognition of the image segmentation algorithm,this thesis designs the overlapped region cropping method.This method solves the problem of false recognition by making the small image boundary contain some adjacent image pixel values.Because of the large scale of the image,the cutting speed is slow.The thesis designs the image cache strategy to improve the cutting efficiency.Secondly,fuse the small images which are segmented by image segmentation algorithm.Because the edge of feature information cannot be marked accurately by image segmentation algorithm,the stitching gap appears when fuse images directly.So this thesis designs the image pyramid fusion algorithm.The algorithm combines the lowfrequency information of the image obtained by Gauss pyramid and the high-frequency information of the image obtained by Laplace pyramid to solve the problem of stitching gaps.Finally,all the small images after semantic segmentation are spliced into a complete semantic segmentation effect image.Because of the number of small images is too big,it is impossible to input them into memory for processing simultaneously.So the thesis designs a batch splicing fusion algorithm.The algorithm splices all small images in batches,and then writes the splicing results into the pre created image file in these batches.Because of the large number of small images,the splicing speed is slow.The thesis designs a memory management mechanism and a multithreading mechanism to improve splicing efficiency.Experiments and practical tests show that the method proposed in this thesis can not only improve the effect of large-scale satellite image semantic segmentation,but also meet the requirements of the actual system for processing speed.