一种地基云图分类算法及硬件加速实现  

Hardware acceleration implementation of a ground-based cloud image classification algorithm

作　　者：冯琳 宋文强 徐伟[1,2] Feng Lin;Song Wenqiang;Xu Wei(Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters Nanjing University of Information Science and Technology,Nanjing 210044,China;Jiangsu Key Laboratory of Meteorological Observation and Information Processing,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学气象灾害预报预警与评估协同创新中心,南京210044 [2]南京信息工程大学江苏省气象探测与信息处理重点实验室,南京210044

出　　处：《电子测量与仪器学报》2025年第2期21-31,共11页Journal of Electronic Measurement and Instrumentation

基　　金：国家重点研发计划政府间/港澳台重点专项(2021YFE0105500);江苏省研究生科研与实践创新计划项目(SJCX24_0470)资助。

摘　　要：地基云的自动观测和识别对分析大气运动趋势和天气预测具有指导意义。针对目前地基云图分类算法准确率不高、在嵌入式终端部署困难的问题,提出了一种基于残差网络结构的地基云图分类网络模型GBcNet及基于ZYNQ的硬件实现架构,PS端用于加载模型的权重参数和云图数据,PL端实现DDR3读写控制和GBcNet的硬件加速。设计了滑窗、卷积层、池化层、批量归一化层和全连接层等模块的加速IP核。实验在CCSN数据集上进行,结果表明,提出的模型在PC端的准确率达到96.02%。采用现场可编程门阵列(FPGA)硬件加速后,准确率仍然保持在94.5%。与PC端模型的识别率相比,各云类的识别精度损失均不超过3%,整体精度损失小于1.5%;FPGA的最大资源占用不超过48%,单张地基云图推理时间为0.13 s。相较于现有地基云的识别方法,识别准确率高且推理时间较短。提出的识别模型和硬件加速方法为便携式地基云观测设备的研制提供了一种参考方案。The automatic observation and recognition of ground-based clouds have guiding significance for analyzing atmospheric motion trends and weather forecasting.To solve the problem of low accuracy of ground-based cloud image classification algorithms and difficulties in deploying them on embedded terminals,a ground-based cloud image classification network model GBcNet based on residual network structure and a hardware implementation architecture based on ZYNQ are proposed.The PS end is used to load the weight parameters and cloud data of the model,and the PL end implements DDR3 read-write control and GBcNet hardware acceleration.For the GBcNet network,accelerated IP cores corresponding to each module were designed,including sliding window,convolutional layer,pooling layer,batch normalization layer,and fully connected layer.Experiments were conducted on the CCSN dataset,and the results showed that the GBcNet model achieved an accuracy of 96.02%on a PC platform.After hardware acceleration,the accuracy remained at 94.5%.Compared with the recognition rate of the PC model,the accuracy loss for each cloud class did not exceed 3%,and the overall accuracy loss was less than 1.5%.The maximum resource usage on the FPGA did not exceed 48%,and the inference time for a single ground-based cloud image was 0.13 seconds.Compared to existing ground-based cloud recognition methods,this approach demonstrates higher accuracy and shorter inference time.The proposed recognition model and acceleration method provide a reference solution for the development of multi-node,portable ground-based cloud observation equipment.

关 键 词：图像分类 地基云图 硬件加速 FPGA 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN919.8[自动化与计算机技术—计算机科学与技术]