Research On Image Change Detection Methods And Applications Based On Difference Feature Learning And Attention Mechanism

Image change detection is a process of comparing and analyzing the images obtained at different periods of the same position,extracting and analyzing the difference features,and finally identifying the changed regions.In recent years,image change detection technology has been widely used in military,agriculture,environmental monitoring,industrial assessment,and other fields.Based on the rapid development of remote sensing technology,image change detection has become a cutting-edge technology in the study of global surface change,which can help people better understand the dynamic changes of the Earth surface.At the same time,the application of image change detection method to oil displacement path detection can provide data support for evaluating displacement agent performance and reservoir simulation technology.Although a large number of researches on image change detection have emerged at the present stage,the following challenges are still faced in practical applications:(1)The current change detection methods have insufficient representation of difference information in the process of feature extraction,which makes it difficult for the algorithm to effectively distinguish pseudo-changed features and cope with change detection tasks of complex scenes;(2)The current change detection methods have the problem that the model complexity is high and it is difficult to deploy in practical tasks.In view of the above problems,this dissertation studies the research on image change detection methods and applications based on difference feature learning and attention mechanism,so as to provide technical support for remote sensing earth monitoring and reservoir displacement evaluation.The main work and contributions of this dissertation are summarized as follows:(1)Image change detection method based on difference feature learning and kernel scale adaptive attentionAiming at the problem that the traditional attention mechanisms neglect the scale difference of different feature maps and limit the spatial flexibility of attention mechanisms,difference feature learning and kernel scale adaptive attention image change detection methods are proposed.Firstly,the relationship between feature mapping and convolution kernel scale is used to design a scale adaptive attention module to obtain the difference feature information effectively.Secondly,this method proposes a multi-layer perceptron module based on image patch embedding,which learns local and global pixel association through skip connection and realizes deep fusion of low-level and high-level features to narrow the semantic gap between high-level and low-level features.Compared to ordinary attention module,this module has stronger ability of feature representation global feature learning.The experimental results on remote sensing image datasets and oil displacement images show that the proposed change detection method has good detection performance,the model parameters is 13.10MB,the model size is 37.13MB,corresponding to the average detection accuracy of 90.25%,and can provide a changed regions with more accurate contours.At the same time,ablation experimental results and analysis further verified the effectiveness and applicability of the proposed change detection method on remote sensing image datasets and oil displacement images,providing technical support for improving the performance of ground monitoring methods and the perfection and application of displacement theory.(2)Deeply supervised image change detection method based on Transformer feature learning and information indexIn view of the difficulty of considering feature spatial information in traditional change detection methods,a deeply supervised image change detection method based on Transformer feature learning and information index is proposed.In this method,feature index information is retained while feature learning is carried out in the encoding stage,and Transformer is combined with multi-head attention mechanism to learn the long-range dependence of changed information.Secondly,in the decoding stage,the advanced semantic feature information of the changed features is learned from the perspective of multi-scale aggregate feature maps by using the deeply-supervised strategy,and the feature index information is used to restore the feature maps with different resolutions to the larger resolutions in the decoding stage.Finally,the change mapping results are generated.Experimental results and ablation results show that the proposed method can learn the long-range dependence of the difference features,and improve the problem of insufficient spatial position features caused by mapping low resolution features directly to the original image resolution.Compared to the comparative methods of the same category,the model parameters and model size are in the middle,which are 41.85MB and 146.92MB,respectively.At the same time,the average detection accuracy of the datasets in this dissertation is increased to 90.39%,and the average accuracy is increased by 1.09%than the optimal comparative method.Finally,this dissertation verifies the effectiveness of the proposed change detection method on remote sensing image datasets and oil displacement images.It provides experience and data support for ground monitoring and evaluation of displacement agent performance.(3)Cross-scale difference feature information image change detection method based on Siamese Swin-TransformerAiming at the problem that current change detection methods has high complexity,and it is difficult to make full use of bi-temporal images features.A cross-scale difference feature information change detection method based on Siamese Swin-Transformer is proposed.Firstly,the method uses the Siamese network of weights sharing to extract the bi-temporal features respectively,so as to better capture the changed features of different scales and levels.Secondly,in the feature extraction stage,a cross-scale difference feature attention module is designed.This module can obtain attention information by convolution operation of the feature information of the current layer and the feature information of the previous layer,and carry out information fusion.Finally,the multi-scale feature aggregation decoding module is combined to achieve the final change detection result.This method has a strong ability of network feature representation,and can make the feature information of adjacent layers guide each other to obtain richer difference features.The experimental results and ablation studies show that the proposed method has good detection accuracy and visual effects on image change detection task.The model parameters and model sizes are 21.57MB and 53.89MB,respectively.The average detection accuracy on the relevant datasets in this dissertation is increased to 90.51%.The average accuracy has been improved by 1.21%.At the same time,the effectiveness of the method on remote sensing image datasets and oil displacement images are verified,which lays a theoretical foundation for ground monitoring technology,displacement agent performance evaluation and the exploration of fluid flow characteristics.(4)Full-scale Swin-Transformer image change detection method based on joint multi-frequency featuresAiming at the problems of high complexity of current change detection methods,difficult deployment,unclear details of detection results and blurry boundaries,a full-scale Swin-Transformer image change detection method based on joint multi-frequency features is proposed.Firstly,the method obtains more comprehensive bi-temporal feature information by designing multi-frequency channel attention module,and improves the ability of modeling channel correlation of feature maps.Secondly,based on the practical requirements of multi-spectral attention module and change detection tasks,a joint multi-multispectral difference feature enhancement guiding block is constructed to realize difference feature learning.In addition,different from the Siamese-based approaches,a full-scale Swin-Transformer module is designed as the third extraction branch,and the pyramid decoder module is combined to model the long-range dependency of multi-scale changed objects.This method makes up for the lack of feature representation caused by only performing global average pooling operation,and enhances the position awareness ability of the model to the real changed objects.The experimental results show that this method can effectively overcome the problem of missing detections of small targets and achieve more complete and compact changed results.The average detection accuracy of the proposed method on the relevant datasets is improved to 91.81%,and the model complexity is smaller.The model parameters and model size are 11.13MB and 43.54MB,respectively.At the same time,the validity and applicability of this method in remote sensing image data set and displacement image library are verified.It can provide technical support for in-depth research on the development of ground monitoring technology and displacement mechanism.