Image-oriented Distributed Parallel Processing System

As satellite imaging technology matures,the application of artificial intelligence technology to recognize satellite remote sensing images has become a research hotspot.However,satellite remote sensing image has the characteristic of high imaging resolution,large format size and high real-time requirement.Therefore,there are some problems such as slow imaging speed and slow recognition speed.When the image recognition speed is much slower than the image imaging speed,the image data will be lost.On the contrary,satellite resources are wasted.Therefore,how to improve the utilization of satellite resources while ensuring the imaging quality of satellite images and improving the recognition speed has become the subject of this thesis.This thesis is designed for an image-oriented distributed parallel processing system.The theoretical basis,design scheme,implementation process and test analysis of the system are elaborated in the thesis.The main work of this thesis is as follows:1)Satellite image fusion on heterogeneous platforms: Satellite image fusion is performed on CPU and GPU on heterogeneous platforms.The same image fusion method is implemented by serial programming and parallel programming respectively.Parallel image fusion method is implemented by CUDA.Choose a more suitable satellite image fusion acceleration method by comparing the time overhead of the two ways at different data sizes.2)Adaptive Weighted Threshold Scheduling Algorithm: An adaptive weighted threshold scheduling algorithm is designed and implemented,which adjusts the image data processing memory and image data transmission memory.The algorithm is divided into four states: initialization,constant,high load,and low load.After starting the task,the module enters the initialization state at first.When the image transmission speed is faster than the image processing speed,the module enters a high load state,and the capacity of the transmission queue will be increased if the memory is sufficient.When the image transmission speed is slower than the image processing speed,the module enters a low load state and the transmission queue capacity will be reduced.The scheduling algorithm strives to improve the utilization of satellite resources and balance the memory usage of the image transmission and image processing.3)High-concurrency low-redundancy storage module: Aiming at the space-time characteristics of satellite image data,a high-concurrency low-redundancy storage module based on LSM-Tree is designed and implemented.The module stores data without central node,and compresses the key value of the picture with the shared prefix compression algorithm.At the same time,the module has index information at the top of each data block to speed up data query.Finally,the module implements two types of garbage collection algorithms to improve the utilization of satellite storage resources.One is a periodic deletion policy that is initiated based on the disk remaining space threshold,and the other is a deletion after querying policy that is initiated after each query.Functional and performance tests on the entire system are performed in this thesis.The test result shows that the satellite image fusion is accelerated;the memory usage for data transmission and data processing are automatically controlled;and image data is distributed stored in the storage module.Storage module successfully improves disk utilization.