Implementation And Verification Of CNN Based On FPGA

With the development of space remote sensing technology,a large number of high-precision and high-resolution remote sensing images are widely used in civil economic construction and military defense.Real-time processing of targets of interest in remote sensing images is a current issue of concern to researchers.In the field of image processing,computer vision technology has always had unique advantages.Convolutional Neural Network(CNN)is a new algorithm in the field of computer vision in recent years.When using CNN to process a large number of remote sensing images in real time,its advantages in the process of feature learning and feature selection,combined with the characteristics of local connection and parameter sharing unique within the algorithm,can quickly extract and identify features in remote sensing images.CNN can not only achieve higher accuracy in image processing,but also can solve the problem of image displacement and deformation.Therefore,CNN can be applied to the remote sensing satellite image real-time processing system.Based on the background of engineering applications,this paper analyzes and summarizes the convolutional neural network algorithm's hierarchical structure characteristics,makes full use of the advantages of Field Programmable Gate Array(FPGA)to build a hardware implementation technology platform,and proposes an FPGA-based platform Convolutional neural network for real-time processing of remote sensing image solutions.The research contents of this article include:(1)According to the project requirements,combined with the structure and characteristics of the convolutional neural network,optimize the LeNet-5 convolutional neural network.By increasing the input image size and increasing the number of convolution kernels,the network is adapted to remote sensing images with a large amount of data;by taking a step of 2 after the convolution layer,the network's intermediate data volume is reduced and the efficiency of the network is improved;By modularizing the network hierarchy,the CNN is adapted to the FPGA hardware platform.(2)Calculate the timeliness of network estimates in conjunction with network-level parameters.By analyzing the parallelism and estimated timeliness of CNN,the FPGA implementation process is divided into two stages of pipelines and started at the same time,and the data is processed in parallel,which improves the efficiency of the system.By analyzing the system requirements and CNN structure,the overall architecture of the realization of CNN software and hardware design by FPGA is designed,and the remote sensing image is identified in real time.(3)According to the overall design requirements of the system,the hardware code is designed in the following three aspects: First,the maximum parallelism of the system is obtained through calculation,and the calculation unit is designed according to the parallelism,so that the modules achieve spatial parallelism and make full use of the bandwidth of the external memory.The two-stage pipeline intermediate data is ping-pong cached to achieve time parallelism and improve network efficiency.Secondly,the module is configured to select different modes of the module to realize module reuse.The data is fixed-pointed and the data cache module is added to the module to make full use of on-chip resources to achieve data reuse.Finally,different-size data storage is used to cache the data between layers,so that the data can be read quickly,reducing Access to off-chip memory.(4)Use the testbench verification method to verify the function of the system.By building a verification platform and combining software to generate incentives,perform functional verification of the design system in this paper.The test verification results show that the overall function of the FPGA hardware system and the functions of each module designed in this paper are correct,and the actual timeliness meets the expected timeliness requirements;the network recognition rate is 97.8%,which has a high accuracy rate.