Change Detection In SAR Image Based On Dilated Convolution And FPGA

Remote sensing image change detection refers to the process of comparing and analyzing the surface information according to the remote sensing image formed at different time in the same region,determining the type of image change information,and quantitatively describing the change.Remote sensing image change detection has important applications in urban changing,land planning,crop growth monitoring,ecological environment monitoring,disaster monitoring and even battlefield surveillance.It is a work of great practical significance for us to detect changes in remote sensing image.Because of the feature of being able to work during all-day-in all-weather condition with penetrating power and the ability to negelect the influence from clouds and sunlight,Synthetic Aperture Radar image change detection has gained more and more popularity.Due to the access of large amounts of SAR data,the fast and accurate SAR image change detection method is the key to the remote sensing data application.To realize fast and accurate change detection of SAR images,this paper mainly studies the key techniques of SAR image change detection based on Dilated Convolution and FPGA.The main research content is as follows: 1.A SAR image change detection method based on Dilated Convolution and balanced loss function is proposed.The method is based on the theory of image segmentation.The ESP space pyramid Dilated Convolution fusion module and the regularized and weight loss function are introduced into the U-Net network.In order to study the detail information of the change region,the number of the pooling layers for the feature in the networks is deduced for less-sampling of the network characteristics.The regularized weight loss function is able to perform network learning under the condition that the numbers of changed and unchanging class samples are not balanced.In this paper,experiments are carried out on multiple sets of change detection data sets.Based on dilated convolution and balanced loss function,the experimental results show that the SAR image change detection method has higher detection accuracy and lower false alarms and false negatives than the traditional clustering or convolution network method.2.Based on OpenCL,a FPGA parallel acceleration system is proposed for SAR image change detection.The system is parallelized and accelerated on the foundation of wavelet fusion and FLICM clustering SAR image change detection method.Firstly,the structure of the algorithm is analyzed,and the parallelization algorithm is designed.Then the working group size,memory access mode and loop expansion are optimized on the FPGA.The algorithm parallelization and optimization greatly improves the speed of SAR image change detection,and the acceleration ratio is more than three times.