SAR Image Change Detection Based On Convolutional Neural Network And Hardware Implementation

Synthetic Aperture Radar(SAR)features high resolution,all-weather,night-and-day functionality,and plays an important role in many applications such as land monitoring and military fields.SAR image change detection is a very important research direction in the field of SAR image processing,which is widely used in disaster detection,urban planning and other fields.At present,there are many studies about SAR image change detection.In this paper,we propose two improved SAR image change detection algorithms based on 3Dconvolutional neural network(3D-CNN).At the same time,in order to improve the speed of traditional detection algorithm,the traditional algorithm of SAR change detection is implemented on FPGA through OpenCL in a parallel way.The main work of this paper is as follows: 1.A SAR image change detection method based on spatial and multi-information map 3D residual convolution neural network(SMIRN)is proposed.In order to solve the problem that it is difficult to fully express the difference information for a single difference map,the multi-difference maps and the original maps are stacked into multi-information maps.The 3D-convolutional neural network is used to extract the features of the multi-information maps.At the same time,a residual module is introduced into the network to prevent the gradient from disappearing and to improve the detection accuracy.The pre-classification result of the image is obtained by two-time ensemble clustering of multi-difference maps.The reliable samples are selected as training samples of the network,and finally the SAR image change detection is realized.Compared with the traditional change detection method and using the 3D-convolution method network directly,the change detection performance of SMIRN is significantly improved.2.Aiming at the problem of multi-scale sample training,on the basis of SMIRN,a method of SAR image change detection based on multi-scale spatial and multi-information maps 3D residual convolution neural network(MSMIRN)is proposed.This method removes the last average pooling layer of the network,and adds a spatial pyramid pooling layer after the multi-information map residual module and the spatial residual module of the SMIRN network.In order to re-calibrate the importance of feature map,MSMIRN adds an SEnet sub-network in the spatial residual module.The MSMIRN method can use multi-scale samples for training directly,and extracts the features of multi-scale samples.Compared with SMIRN,the detection accuracy of MSMIRN is improved.3.A method of SAR image change detection based on OpenCL parallel acceleration is proposed.This method adopts multi-difference maps fusion strategy,and then realizes the change detection of SAR image by clustering analysis of the fusion difference map.Aiming at the parallel optimization problem of the algorithm,generating a difference map,mean filter,median filter,fusing a difference map and K-means clustering modules are implemented on the FPGA,and the algorithm is further accelerated by optimization methods such as kernel vectorization.Under the premise of ensuring the accuracy of the algorithm,the acceleration of the change detection algorithm is realized.