| With the rapid development of remote sensing technology,it has become more and more closely integrated with artificial intelligence technology,and has gradually penetrated into all aspects of the remote sensing field.Nowadays,the resolution of remote sensing images is gradually improving,how to efficiently use high-resolution remote sensing image data to automatically extract multi-scale target feature information as quickly as possible and reduce the loss of target feature information has become a hot spot in the field of remote sensing research.Among them,roads and buildings are both important components of geographic information and are commonly used in urban planning,intelligent transportation,land use and other fields.Road features usually have linear scale transformations,multiple road intersections and other features that make extraction difficult;the building features usually have the complex situation of sparse or dense distribution,and the combination of buildings of different sizes and orientations makes it difficult to extract.For the different problems faced by key target features in high-resolution remote sensing image scenes,this paper proposes different solutions for road and building target extraction respectively based on the learning theory of multi-scale feature fusion,and the main work is as follows:(1)To address the problem of road,a key feature information in high-resolution remote sensing images,the lack of linear features in the extraction process leads to incoherent roads and incorrect recognition at intersections,etc.The overall effect of road extraction is improved by designing a multi-scale feature fusion network model(DC-ASPP-Res Unet).On the one hand,the perceptual field of feature extraction is further expanded by hybrid null convolution based on the network design of depth residuals,thus making the semantic information of shallow roads in the multi-scale feature fusion module richer;on the other hand,the accuracy of obtaining shallow road semantic information in multi-scale fusion features is improved by using weight distribution matching sensory field scales,and the improved porous space pyramid pooling module is integrated into the deep residual network to achieve deep aggregation of road information at different scales.(2)To address the problem that buildings,another key feature information in highresolution remote sensing images,are more prone to category imbalance in the extraction process than road features,leading to difficulties in extracting different size building targets and poor adaptation of multi-scale features,the overall effect of building extraction is improved by designing a dual multi-scale attention mechanism feature fusion network model(DMAF-RASnet).First,the coding side of the residual network is constructed using the channel attention module to enhance feature information extraction,and the porous space pyramid pooling module is incorporated between coding and decoding to extract multi-scale feature information;second,the correlations and global dependencies between building features are integrated at the decoding end using channel grouping mash and double attention mechanism fusion;finally,considering the complex and variable nature of building data,the hybrid loss function is designed to handle the imbalance of categories,which makes the network model training more stable.In this paper,we conduct an in-depth study on the multi-scale feature fusion problem in target extraction methods,and design different multi-scale feature fusion methods in the process of road extraction and building extraction tasks,respectively,to solve the problems faced in different key feature extraction tasks.At the same time,both methods are able to have substantial improvement in the accuracy of the experimental results,highlighting the efficiency of the method,which is of great importance in the further study of the target extraction network model in the future as well as in the practical application of mining out the related methods. |
PDF Full Text Request