Research On Change Detection Technology Based On Feature Extraction

Change detection is an effective means of automatically detecting surface changes and plays an important role in many practical applications.With the improvement of spatial resolution of remote sensing technology images,remote sensing image features become more and more complex,and some early change detection methods are difficult to meet the demand.Analysis of remote sensing image change information using traditional methods is time-consuming and laborious,making its automation an important area of research.In recent years,computer vision algorithms based on deep learning have been fully developed,especially in the field of change detection deep convolutional neural networks have achieved good results.In fact,change detection is a binary classification problem for semantic segmentation.Therefore,in this paper,the task of change detection is performed on the basis of semantic segmentation networks,and two different convolutional neural networks are used to perform change detection on remote sensing images.One is a deep supervised network with end-to-end multi-side output fusion to accomplish the research task in this paper;the other is a remote sensing image change detection network with twin high-resolution representation to accomplish the change detection task.The research of this paper is mainly as follows:For different input methods of change detection networks,this paper combines images from two time periods into a single image input network for training,a deep supervised network with end-to-end multi-side output fusion is proposed.The network increases the perceptual field and captures multi-scale features by enhancing feature extraction module.To balance the low-level spatial detail information of the shallow network and the high-level abstract semantic information of the deep network,the dense skip connection module is introduced.In addition,a deep supervision module with multiple side-outputs fusion is introduced to improve the robustness of object scale variation and facilitate the detection of small targets.Experimental results on the publicly available CDD dataset of high-resolution remote sensing images and a challenging DSIFN dataset demonstrate that the network proposed in this paper achieves much better change detection performance.Compared to the ENCLNet,where relatively high experimental results are obtained so far,the F1 score improved by 1% and 5.53% on the two datasets,respectively.When performing change detection,the merging of image channels may cause distortion of the image due to the lack of detail and feature compactness in the experimental results.A remote sensing image change detection network based on twin high-resolution representation is designed and implemented.In order to reduce the loss of spatial accuracy of remote sensing images and maintain a high-resolution representation,HRNet is migrated as the backbone network for initial extraction of image features.The context aggregation module is introduced to obtain not only global information but also more detailed local contextual information.In order to improve the accuracy of small volume target change detection,while shortening the training process and speeding up the detection,the side output embedding module is proposed.Compared with the ENCLNet network,this scheme improves the F1 scores by 0.55%,6.71%and 2.26% on the CDD,DSIFN and SYSU-CD datasets,respectively.