Research On Homogeneous And Heterogeneous Remote Sensing Image Change Detection Based On Capsule Network Learning

This paper studies the change detection in remote sensing images mainly.And this research on change detection has two research directions,homogeneous image change detection and heterogeneous image change detection.The remote sensing image change detection technology is mainly used to detect the same and different regions of the two images acquired by the satellite at different times in the same place.With the development of remote sensing technology and satellite technology,remote sensing images are more and more easy to obtain.We can find out the development and changes of a certain city,the increase or decrease of wetland area and quickly perform in earthquake or fire rescue based on the large number of remote sensing images obtained.Images with different feature spaces,multi-channel,high resolution,etc.are more and more popular among researchers.And many related technologies have been greatly developed.At the same time,with the development of computer technology,deep neural network technology has made great progress.And deep neural networks have been applied in many aspects,especially in the field of image research.The previous part of the study is the detection of changes in the image of Synthetic Aperture Radar(SAR).The latter part is the detection of the change of heterogeneous remote sensing images,mainly based on the study of SAR image and optical image.It includes three specific studies mainly:A homogeneous remote sensing image change detection method based on capsule network and binary image fusion is proposed.The method first uses the log-ratio(LR)method to calculate the logarithmic ratio of two original images to obtain a rough difference image,and then uses a fuzzy clustering method Fuzzy C-means(FCM)to classify the difference image to obtain the initial the binary classification map.Finally,the training samples are selected according to two different size image blocks,and then the images are input into the capsule network for training.The two detection results are then used to perform binary image fusion,and the final change detection result is obtained.A homogeneous remote sensing image change detection method based on Kittler and Illingworth(KI)threshold method and deep capsule network is proposed.The method first calculates the pixel similarity of two original experimental images,obtains the similarity threshold,and then pre-classifies the two original experimental images by using the KI threshold method,and then compares them with the classified images and the similarity threshold toobtain a rough classification result at each pixel.The training samples are then selected based on the ratio of the same as the central pixel classification result in the image block of the classification result map.The capsule network is deepened,the obtained advanced feature information is classified,and finally the final change detection result map is obtained.A heterogeneous remote sensing image change detection method based on Self-organizing feature map(SOM),image mapping method and deep capsule network is proposed.The method firstly uses the Self-organizing feature map network to roughly identify two original experimental images to obtain a changed region and an unchanged region.Then in the unchanged area,some small image blocks are selected for image mapping.The image mapping method is based on these unchanged small image blocks,and the image pixels located in different feature spaces are respectively transformed.The converted pixel has the same or similar feature space information as another original experimental image.Thus,the converted image can be directly compared with another original image.Two original experimental images are converted separately,and then compared with another original experimental image,obtaining two difference images finally.The two difference images are fused to obtain a difference image for selecting samples and training the network.The samples selected from the two images are directly connected and input into the deep capsule network to obtain the final change detection result.