Research On Semantic Segmentation Algorithm Of High Resolution Remote Sensing Image Based On Codec Structure

At present,remote sensing images gradually show the characteristics of high temporal resolution,high spectral resolution,and high spatial resolution.The rich ground object information reflected by remote sensing images is a valuable resource in various fields.Most of the existing remote sensing image segmentation methods are based on deep learning semantic segmentation methods,but these methods still have some shortcomings.On the one hand,in order to solve the problems of incomplete targets and blurred edges in the segmentation effect of existing algorithms for high-resolution remote sensing images,a multilevel feature aggregation network based on the current popular encoder-decoder structure is proposed.It mainly uses high-level features to provide partial semantic information for lowlevel features to guide their positioning through multi-level feature attention up-sampling modules that form a decoder,which effectively recovers the detailed information of highresolution remote sensing images.On the other hand,due to the limitation of short-range contextual information,the current remote sensing image semantic segmentation algorithms cannot fully recover high-resolution detail information,especially edge information.To this end,a multi-level aggregation network based on global dependencies is proposed.It adds a global dependency module in the middle of the encoder and decoder to calculate the affinity matrix between all positions to obtain the important global context information,and adds a two-way feature extraction to assist the up-sampling module to reorganize the latest features.The method further effectively enhances the model's abilities to extract features and recover high-resolution details.Experimental results show that the two algorithms proposed in the paper achieve satisfactory segmentation performance for high-resolution remote sensing image datasets.