Research On Real-time Cloud Detection Of Remote Sensing Image With Convolutional Neural Network

In the visible light remote sensing technology,the widespread cloud cover phenomenon will significantly reduce the utilization value of the image,resulting in a waste of transmission bandwidth and storage space in the on-borad computing platform.Therefore,preprocess images on orbit based on cloud detection technology,eliminate invalid images occluded by clouds,and improve the utilization efficiency of remote sensing images is a hot spot in the field of remote sensing image on-orbit processing.In recent years,remote sensing image cloud detection methods have gradually developed from traditional methods such as threshold method and texture analysis to artificial intelligence methods with convolutional neural networks.The cloud detection with convolutional neural network has the advantages of high detection accuracy and strong robustness,but its huge amount of parameters and calculations make it difficult for the spaceborne platform based on the general-purpose processor of the simplified instruction set architecture to match its computing performance and resource requirements,which limits its real-time processing applications on orbit.In order to achieve real-time high-precision cloud detection under space-borne conditions with limited power consumption and storage resource,a customized energy-efficient processor platform is required to deploy and accelerate cloud detection algorithms with convolutional neural networks.In order to solve this problem,this thesis proposes an FPGA-based lightweight cloud detection convolutional neural network construction and accelerated calculation method.First,based on U-Net,which is widely used in the field of remote sensing image semantic segmentation,a lightweight semantic segmentation cloud detection network based on encoding-decoding structure is proposed,which can effectively reduce the amount of model parameters while ensuring detection accuracy;then,in order to improve calculation efficiency and accelerate inference speed of the network,parameter pruning method based on the channel scaling factor of the BN layer and the two-dimensional Winograd algorithm are used to compress the network and accelerate the calculation,and combine the characteristics of FPGA parallel computing to customize the high-efficiency calculation acceleration unit;finally,based on the "ARM+FPGA" heterogeneous processor,the cloud detection computing processing system is designed and the complete image data input and detection result output path is constructed.This thesis attempts to build a lightweight semantic segmentation network for cloud detection,reducing the amount of model parameters while ensuring a certain detection accuracy.This thesis is also aimed to reduce the computational complexity of the network through model compression and hardware acceleration methods and further improve the speed of cloud detection.The network will be deployed in Xilinx’s heterogeneous computing system platform to verify the feasibility and real-time performance of the cloud detection method in orbit.