Change Detection In SAR Images Based On Deep Capsule Network And Model Compression

Synthetic Aperture Radar(SAR)image is an image generated by the active emission of electromagnetic waves from a SAR sensor and the reception of echoes.Because of SAR's characteristics of strong penetration,long distance and all-weather and all-day,it can obtain high-resolution SAR images without the influence of light intensity,cloud cover and weather.SAR image change detection is a technique to find out the changed part and the unchanged part by observing the SAR image of the same object in the same place at different times.SAR image change detection has drawn increasing attention from researchers because of its applications in agricultural detection,urban planning,disaster assessment and forest early warning.However,due to the scattering phenomenon,the difficulty in SAR image processing is speckle noise,which seriously affects the spatial information of the image and brings difficulties to the subsequent change detection and other tasks.In recent years,the deep neural network has made a breakthrough in the field of image recognition.Therefore,this paper combines the deep neural network with the problem of SAR image change detection,and combines the problem of high model complexity and parameter redundancy faced by deep learning,and applies the deep learning model compression to the SAR image change detection.The main points of this paper are as follows:1.Aiming at the negative impact of speckle noise in SAR image and the difference map generation operator on the original image information in the change detection,this paper proposes a SAR image change detection method based on the deep capsule network.This method directly takes the image block in the original image as the input of the network,and firstly extracts the features in the original image through the convolution layer.After that,the features are integrated and routed by capsule neurons.The features in the capsule layer are no longer a single number in the convolutional neural network,but a vector representing the features,which contains more abundant image feature information.The coupling coefficient between capsule neurons was updated by dynamic routing algorithm.This method has a good inhibitory effect on speckle noise in SAR images.Moreover,compared with the general supervised SAR image change detection method,the capsule network requires less training samples to achieve better results due to its better feature extraction and feature preservation capability.Experiments using the model on fourdifferent sets of standard data have proved that the model has good detection performance.2.Aiming at the problems of high complexity and large number of parameters of deep learning model,this paper proposes a compression method of SAR image change detection model based on weight pruning.Based on the proposed deep capsule network,the parameters are analyzed layer by layer,and the pruning threshold of the layer is determined by pruning ratio.Subsequently,the pruned network was fine-tuned to improve its detection performance.Finally,the storage space occupied by the model is compressed by compressing and storing the parameters left in the model.Experiments on multiple real SAR images demonstrate the feasibility of the proposed model compression method for SAR images and deep capsule network.