Semantic Segmentation Of High Resolution Remote Sensing Images

High-resolution remote sensing images often contain complex semantic information and confusing targets.It is an important and challenging semantic segmentation task for high-resolution remote sensing images.Based on DeepLab V3+or Deep U-Net network structure and tree-like neural network structure module,we propose a semantic segmentation network for high-resolution remote sensing images.The proposed network structure not only modified DeepLab V3+structure to make it applicable for multi-scale and multi-modal data,but also added and concatenated a tree-like neural network block.The tree-like shape structure is constructed by establishing a confusion matrix,extracting the confusion graph,and graph cutting operation,which can segment the easy confusing pixels better and obtain more accurate segmentation results.The experimental results in two different city datasets provided by the ISPRS committee show that the proposed network jointly using segmentation structure and Tree-CNN block can effectively improve overall semantic segmentation accuracy of high-resolution remote sensing images.Among them,the segmentation accuracy of easy confusing category is improved significantly.The experimental results show that in high-resolution remote sensing images with complex semantic information,the overall segmentation accuracy of the proposed network with the tree-like block is greatly improved due to the reduction of the error on easy confusing pixels.The proposed method in this paper is universal and has a wide range of application scenarios.