SAR Image Change Detection Based On Adaptive Self-Paced Learning

SAR image change detection is to calculate the area of change from two SAR images of the same area at the different time.How to suppress the influence of various noises in SAR image and get more accurate result of change detection is an important subject of research now.Self-paced learning is to simulate the human learning process,from simple to complex gradually learning samples,the method can avoid the model of learning into a local optimum and suppress the influence of the noise effectively,but has the problem of parameters is difficult to adjust.This paper studied the self-paced learning based on particle swarm optimization.Then aiming at the shortcomings of the existing SAR image change detection,SAR image change detection based on adaptive self-paced learning is proposed.The paper mainly completed the following three tasks:For the problem that the age parameter of self-paced learning is difficult to adjust,an adaptive self-paced learning based on particle swarm optimization is proposed.The particle swarm optimization algorithm is used to optimize the age parameter so that the age parameter can be adjusted adaptively.By analyzing the characteristics of age parameters,a new coding method is proposed to encode the age parameters,reduce the searchable feasible space,and improve the convergence speed of the algorithm.A fitness function for evaluating the age parameter coding is designed by analyzing the objective function of self-paced learning.On the two experiments of matrix factorization and motion recognition,the proposed algorithm effectively solved the problem of adjusting the age parameters,can find better age parameters at the same time.The learning effect of the proposed algorithm is better than the traditional self-paced learning method.Considering the effect of noise on the change detection results in the existing change detection methods,adaptive self-paced learning is applied to SAR image change detection to improve the ability of suppressing noise.SAR image change detection based on self-paced learning and symmetric convolutional coupling network is proposed.The self-paced learning is used to select the pixels in the SAR image to train the symmetric convolutional coupling network.The network preferentially learns simple pixels,and then gradually learns the complex pixel points,so that the symmetric convolutional coupling network for the SAR image learning is more in line with the path of machine learning and can avoid the network learning being trapped in a local optimum.At the same time,the algorithm is analyzed theoretically,the objective function of the algorithm and the optimization procedure of the objective function are proposed.Experiments show that the proposed algorithm has better effect of noise suppression than existing SAR image change detection algorithms.For the existing SAR-based change detection method based on deep neural network,when training samples are selected,misclassified samples may be selected as training samples,so that the training of the network is misled.In this paper,SAR image change detection based on self-paced learning with diversity and PCANet is proposed.Two SAR images are processed by FCM algorithm to obtain initial change detection results.Training samples are selected from initial results,then are grouped by FCM algorithm.A PCANet neural network was trained using adaptive self-paced learning to obtain the final change detection result.The algorithm preferentially learns simple pixels,and then gradually learns complex pixel points,which can reduce the misleading for neural network training.At the same time,taking into account the diversity of samples,the sample is more reasonable in training.The validity of the proposed algorithm is proved on two SAR image data.