Semantic Segmentation Of Urban Remote Sensing Images Based On Convolutional Neural Networks

The types of land cover within remote sensing images are often highly imbalanced,large object classes dominate the segmentation task,and small object classes are usually suppressed,so class imbalance is one of the serious problems that plagues the semantic segmentation task in urban remote sensing images.In view of this situation,this thesis mainly studies and analyzes the semantic segmentation of small object class in urban remote sensing images,and a method of semantic segmentation based on convolutional neural networks is proposed.The main work of this thesis is as follow:?1?In light of the class imbalance of the semantic segmentation in urban remote sensing images,a focal loss function weighted by the median frequency balancing(MFB_Focal_loss)is given.MFB_Focal_loss weights the class loss by the ratio of the median class frequency in the training set and the actual class frequency.A factor is introduced based on standard cross-entropy loss to inhibit the leading role of large object classes during training and thus focus training on small object classes.Our experiments were based on the Vaihingen dataset and the experimental results show that,in the case where boundary pixels were considered,the MFB_Focal_loss achieved a good overall segmentation performance using the same U-Net model,and the F1-score of the small object class“car”was improved by 9.28%compared with the cross-entropy loss function.It is proved that MFB_Focal_loss can effectively improve the segmentation performance of the small object class.?2?In order to improve the accuracy of urban remote sensing image segmentation,this thesis developed the concept of the Down-sampling Block?DownBlock?for obtaining context information and Up-sampling Block?UpBlock?for restoring the original resolution,and an end-to-end deep convolutional neural network?DenseU-Net?architecture is proposed.The main idea of DenseU-Net is to connect convolutional neural network features through cascade operations and use its symmetrical structure to fuse the detail features in shallow layers and the abstract semantic features in deep layers.The experimental results show that,in the case where boundary pixels were considered,using the same MFB_Focal_loss loss function,the F1-score of the“car”class by using the DenseU-Net is increased by 6.71%compared with by using the U-Net,while the average F1-score and overall accuracy are improved by 3.23%and 2.42%respectively.It is proved that DenseU-Net can effectively improve the segmentation performance of the small object class and overall accuracy.?3?In order to analyze the input images of two different statistical characteristics,the SiameseDenseU-Net is proposed based on DenseU-Net.The SiameseDenseU-Net uses two parallel DenseU-Net to extract deep features of the input images,and then concatenate the features extracted by two parallel DenseU-Net.Finally,the DenseCRF is used as the post-processing method to optimize the segmentation results.The experimental results show that the segmentation performance of the SiameseDenseU-Net still outperforms that of the original DenseU-Net without additional parameters and calculation cost.In the case where boundary pixels were considered,using the same MFB_Focal_loss,the F1-score of the“car”class by using the SiameseDenseU-Net is increased by 0.9%compared with by using the DenseU-Net,while the average F1-score and overall accuracy are improved by 0.57%and 0.56%respectively.