Font Size: a A A

Semantic Segmentation Of High-Resolution Remote Sensing Images Based On Deep Learning

Posted on:2021-05-24Degree:MasterType:Thesis
Country:ChinaCandidate:Z B HuFull Text:PDF
GTID:2392330620468781Subject:Engineering
Abstract/Summary:PDF Full Text Request
High-resolution remote sensing images have become an important window for humans to observe ground objects.Semantic segmentation of remote sensing images completes the classification and recognition of targets by assigning semantic tags to each pixel of the image.It is one of the important means of remote sensing image understanding and is widely used in feature changes,urban changes,disaster relief,etc.The high-resolution remote sensing image has the characteristics of high spatial resolution,complex types of ground features,and large intra-class differences,which leads to lower semantic segmentation accuracy under the traditional method.As deep learning has demonstrated excellent performance in visual tasks such as image classification and target detection,in recent years,there have been more and more studies on image interpretation and visual analysis using deep learning methods.Deep learning breaks through the traditional method that requires manual participation in the design of features,and automatically learns feature information from a large amount of image data,which brings new ideas to the study of remote sensing image semantic segmentation.Based on the convolutional neural network,this paper has carried out related research work on high-resolution remote sensing images.By studying and analyzing the characteristics of deep learning technology and common semantic segmentation models,a network model based on dual attention multi-scale feature fusion is constructed.The model mainly has the following characteristics: For high-resolution remote sensing images with more details,in the coding part,ResNet50 is used to extract features,and in the last two stages of ResNet50,the hole convolution method is used to increase the receptive field while maintaining the same amount of parameters,and to capture more global information;in response to changes in target scales,pyramid pooling module is introduced at each stage of ResNet50 to make full use of multi-scale contextual information;a dual attention module is introduced at the final output feature of feature extraction to spatial and channel dimensions semantic dependency modeling is used to enhance the ability of feature representation;in the decoding part,starting from the output feature of the attention module,the feature information of each level of the coding part is gradually fused to complete the decoding and refine the target segmentation edges.Aiming at the characteristics of remote sensing image with complex features,high resolution,and large feature scale differences,this paper uses CCF remote sensing image data sets to analyze the proposed method.First,the pyramid pooling and dual attention ablation experiments were carried out separately.The results show that the pyramid pooling structure can better segment features with large scale differences,and it performs well on multi-scale targets such as roads and water bodies.The attention module has a certain effect on the segmentation of small target areas such as water bodies and noise interference.The method proposed in this paper is compared with FCN,SegNet,U-Net models.The experimental results show that the method proposed has a good semantic segmentation effect,can completely segment the target,and has less noise.
Keywords/Search Tags:deep learning, high resolution remote sensing image, semantic segmentation, dilated convolution, pyramid pooling, dual attention
PDF Full Text Request
Related items