| SAR image change detection has great significance in many fields,and it has always been one of the research hotspots in the field of remote sensing image interpretation.With the continuous development of SAR technology,the resolution of SAR images has been greatly improved,and the amount of acquired SAR image data has become larger and larger.at the same time,the rapid development of the information age has driven people to have higher and higher requirements for the real-time processing of SAR image.Therefore,under the premise of ensuring the accuracy of SAR image processing,it is especially important to increase the processing speed.With its excellent features such as many cores and high bandwidth,GPUs have unique advantages in image processing.Using the GPU to perform highly threaded parallel processing tasks during SAR image change detection can achieve significant speed performance improvements.The emergence of the CUDA architecture has made the development of GPU programs easier.CUDA enables heterogeneous processing models to work better together between CPUs and GPU and even multi-GPU.In the CUDA architecture,based on the CPU+GPU and CPU+multi-GPU processor heterogeneous framework,the two common SAR image change detection algorithms are decomposed in parallel,and the processing speed is greatly improved.The specific engineering practice results of this thesis are as follows:(1)A SAR image change detection algorithm based on wavelet transform is optimized in parallel.The algorithm performs multi-scale decomposition based on wavelet transform on difference images.By performing Bayesian clustering segmentation on multi-scale components,and then performing scale-to-scale fusion on each scale classification result,the final SAR image change detection result is obtained.In this paper,the algorithm is implemented in parallel under the CUDA architecture based on the CPU+GPU processor model.For the image transposition process that is frequently executed,the image to be transposed is divided into sub-blocks having the same dimension as the thread block of the transposed kernel function,and the sub-block data is transmitted to and from the shared memory through the method of merging and fetching.This method reduces the number of data reads and writes between the GPU processing core and the memory.The convolution process of image data and wavelet kernel is constructed into a mathematical model of multiplication of several matrices.By dividing the matrix into a board,each thread block corresponds to a sub-block in the result matrix,so that data is read and written between shared memory and register in sub-portion,which not only utilizes the high-speed characteristics of the shared memory,but also fully utilizes the locality principle to reduce the access delay.Through comparison experiments,the parallel optimization method implemented in this paper greatly improves the processing speed under the premise of ensuring the accuracy of change detection.(2)A MRF-based SAR image change detection algorithm is also optimized in parallel.The algorithm firstly classifies the difference image based on the FCM algorithm,and then constructs the probability model of the difference image and the initial classification result.The final binary classification result is obtained by performing an ICM-based iterative process on the random field.In this paper,the algorithm is implemented in parallel under the CUDA architecture based on CPU+multi-GPU processor model.The correlation functions of the two clusters in the FCM process are respectively allocated to two GPU calculations,and communication between the two GPUs is performed by zero-copy memory.For the ICM process,by zoning the random field to decompose computation into multiple GPUs,the load balancing is realized and the resource utilization is improved.Through comparison experiments,the parallel optimization method implemented in this paper not only improves the accuracy of change detection,but also greatly improves the efficiency of change detection. |
PDF Full Text Request