Change Detection For High-resolution SAR Images With Salient And Incremental Deep Learning

Image change detection is to analyze two or more images at different times in the same area and obtain the changed areas in the image.Synthetic Aperture Radar(SAR)has the ability to work all-time without the limitations of weather and climatic conditions,making it be widely used in the field of agricultural,forestry,water,natural disasters and military,so that SAR images have become the main source of the data for change detection.This paper studies SAR image change detection based on unsupervised methods and proposes three new change detection algorithms.1.This paper proposes a method of high-resolution SAR image change detection with SAE and saliency.The method includes two parts: extracting saliency areas and detecting changed areas.In the process of extracting saliency areas,firstly,we build and train the stacked auto encoder network and obtain the feature representations of the original image,and then the difference map is obtained by analyzing the feature differences.A threshold is used to segment the difference map to obtain a saliency areas in a specific scale.Then,the network structure and the scale of sample block are modified to obtain the saliency areas at other scales.Finally,the saliency areas at multiple scales are merged to obtain the final saliency areas.In the process of detecting change areas,the final change detection results are obtained by clustering the saliency areas with fuzzy C-means clustering method.The key lies in the introduction of the stacked auto encoder network in the process of extracting saliency region,which overcomes the problems of the use of image implicit information insufficiently and the difficulty of threshold selection in previous saliency detection methods;the fusion of multi-scale saliency areas has increased the recall rate of the region areas.That is more conducive to the improvement of the detection accuracy of the later changes.2.This paper proposes a high-resolution SAR image change detection method based on the curvelet SAE network.This method combines curvelet transform with the stacked auto encoder,and the detection procedure includes two parts: extracting saliency areas and detecting changed areas.In the change area detection,firstly,the multi-scale and multi directions features is extracted by curvelet transform,and then deep features is obtained by training the stack auto encoder network on the basis of the multi-scale and multi direction features.Finally,the change detection result is obtained by clustering the difference of the deep features.By introducing the curvelet transform,the input of the stack auto encoder network has the features of multi-scale and multi-directions,which strengthens the feature representation capability of the network.It can better capture the texture information of the image and facilitate the separation of the change area and the unchanged area.The change areas in texture are detected better.3.This paper proposes a method of high-resolution SAR image change detection based on incremental CAE network.On the one hand,the incremental learning method is applied to the training process of convolutional auto encoder network,and an incremental convolutional auto encoder network(ICAE)is proposed,which improves the generalization performance of the network and the ability of the network to adapt data greatly.On the other hand,the clustering results are used as the supervised information of the ICAE network to tune network finely,making the network to learn the features related to the change detection task,then an optimal state is achieved by optimizing the clustering process and network training iteratively.This method performs better on all of the five data sets.