Research On Sensing Image Change Detection Model Based On Adaptive Weight And Multi-scale Convolutional Neural Network

With the rapid development of aerospace technology,the difficulty of obtaining remote sensing images is gradually reduced,and the tasks of automated understanding of remote sensing images are becoming more and more complex and arduous.Remote sensing image change detection is an important application in the field of remote sensing.With the development of deep learning technology,the use of convolutional neural networks as a model for remote sensing image change detection has become the mainstream trend and has achieved good performances.Existing remote sensing image change detection deep learning models often only discuss the effect of change detection in a data set with a specific resolution.Resolution of remote sensing images used by different remote sensing missions may be different,therefore,discuss the robustness of the remote sensing image change detection model in different resolution data sets has important research significance.Based on the deep learning technology,combined with the change detection algorithm of the convolutional neural network,this paper carried out the research on the remote sensing image change detection model based on adaptive weight and multi-scale convolutional neural network,this paper proposed Weighted Rich-scale Inception Coder Network(WRICNet),which can make a fusion of shallow multi-scale features and deep multi-scale features.The main content and work of this paper are as follows:(1)The Rich-scale block is improved to strengthen the independence of each grouping feature and improve the change detection effect by removing the feature fusion between groups and using a different number of convolution in each group without changing the scale obtained by each group.(2)The Inception module was improved,and the model parameters are reduced without changing the scale obtained by each branch.(3)Proposed a Weighted scale block,which can assign appropriate weights to features of different scales,so that it can more accurately represent the edge of the change area,which is beneficial to reduce the missed and false alarms at the edge of the change area.(4)The proposed method and typical methods CDNet,FC-EF,FC-Siam-conc,FC-Siam-diff and STANet are compared and experimented on three different resolution data sets,the results proved that the method in this paper had strong robustness on multi-resolutions data set.And this paper conducted ablation experiments on proposed method,the results proved the effectiveness of the training strategy in the experiment process and improvements of model.