Superpixel-level Image Change Detection Based On Deep Learning

The purpose of the change detection task is to lock the changing regions through algorithms when the images at different times in the same area are given.The traditional change detection algorithm takes pixel as the basic analysis unit,which gains high performance in dealing with simple scenes or low-resolution scenes.But the performance of the algorithm will be limited when faced with complex scenes or high-resolution problems.In order to solve these problems,superpixel is taken as the basic processing unit of change detection algorithm in this thesis,deep learning is also used for feature learning and classification because of its hierarchical structure.The main contents of this paper include the following three aspects:First,a method of superpixel level change detection based on convolution neural network is proposed for dealing with the problem of multi-temporal remote sensing image change detection.In this method,an algorithm called combined simple linear iterative clustering segmentation algorithm specific to the change detection problem is proposed in this paper,which incorporate with the information of multi-temporal image into segmentation progress for remaining the same segmentation contour.In order to measure the similarity of superpixels,a simple superpixel similarity metric method is proposed,the similarity between superpixels are represented by probability.Convolutional neural network is constructed for classification of changed and unchanged class.Second,a superpixel level change detection method based on deep belief network is proposed for dealing with the ternary change analysis problem.In this method,unsupervised change vector analysis algorithm is used for pre-classification of superpixel.Then,a simple but effective sample selection method is used for selecting pure samples as dataset.At last,deep belief network is established for superpixel classification.For overcoming the problem of inaccurate segmentation,multiscale strategy is also used in this method.Thirdly,for accomplishing heterogeneous images change detection task,a superpixel level heterogeneous change detection algorithm based on coupling convolutional neural network is proposed in this paper.In this method,a new structure and training progress of deep neural network is designed for dealing with heterogeneous images change detection problem,which mapped heterogeneous images into the same feature space simultaneously.During the training progress,a new loss function is designed,which is combined with clustering operation to guide the segmentation of clustering algorithm through the training results of network,and the direction of network parameter adjustment is guided by the segmentation results of clustering algorithm.Theoretical analysis and experiments prove that superpixel level change detection based on deep learning can deal with multi-temporal image change detection problem effectively and accurately.