遥感图像实时检测是遥感应用领域的关键技术问题之一, 针对目前主流的目标检测算法在图像处理器(GPU)上存在模型参数量大、实时性差、功耗大和成本高的问题, 提出一种基于现场可编程门阵列(FPGA)的遥感图像实时检测方案。首先, 为减少参数量、提高检测速度, 采用MobileNetv2作为特征提取网络, 融合深度可分离卷积, 使得模型轻量化便于部署; 接着,采用CA注意力模块提高检测精度;最后, 将模型浮点数参数量化为8位定点数, 并将量化后的网络模型在FPGA上完成部署。实验表明, 在遥感数据集VisDrone 2019上, 所提设计方案平均精度均值(mAP)达到14.79%, FPS达到46.78 帧/s, 平均功耗为8 W, 比CPU提高375.4%的检测速度, 比GPU降低96.8%的功耗。该方案可以满足实时目标检测的要求, 并且能够部署在功耗受限的卫星、无人机等设备上。

Real-time detection of remote sensing images is one of the key technical problems in the field of remote sensing application.As for current mainstream target detection algorithms, there are problems of a large number of model parameters, bad real-time performance, high power consumption and high costs on the image processor (GPU).To solve the problems, a real-time detection scheme of remote sensing images based on Field Programmable Gate Array (FPGA) is proposed.Firstly, in order to reduce the quantity of parameters and improve the detection speed, MobileNetv2 is taken as the feature extraction network, and the fusion with depth separable convolution is conducted, making the model lightweight and easy to deploy.Then, CA attention module is used to improve the detection accuracy.Finally, the floating point parameters of the model are quantified into 8-bit fixed point numbers, and the network model is deployed on FPGA after quantization.The experimental results show that on the remote sensing data set VisDrone 2019, the mean Average Precision (mAP) of the scheme designed in this paper reaches 14.79%, FPS reaches 46.78 frame/s, and average power consumption is 8 W.The detection speed is 375.4% higher than that of CPU, and the power consumption is 96.8% lower than that of GPU.The scheme can meet the requirements of real-time target detection, and can be deployed in power-limited satellite, UAVs and other equipment.