关键词: 星载平台, 遥感图像, 深度学习, 国产器件, 模型部署优化, 舰船目标检测

Abstract:

To address the problems of limited size and power consumption of the on-board computing platform， the hardware architecture of “FPGA+AI processor” is proposed， and the on-board image processing platform centering on the domestic commercial devices Fudan Micro JFM7K325T FPGA and Huawei Ascend310 AI processor is designed. Intelligent computing platform. To overcome the problem of high computational and storage costs for deep network algorithms， a ship target detection algorithm based on the lightweight remote sensing image is proposed， which effectively reduces the amount of computation and parameters while ensuring the detection accuracy. To solve the problem of slow computing in real-time deployment of neural network algorithms， we carry out quantization-aware training， calculation graph optimization of operator fusion， and multi-threaded parallel software design， which effectively improves real-time computing performance. The experimental results show that the power consumption of the system is less than 15 W， the detection accuracy of ships in remote sensing images reaches 0.908， the frame rate of 1 152×1 152 remote sensing image processing reaches 63.3 frames/s， and the frame rate of 4 096×5 120 large-format image processing reaches 3.17 frames/s. This paper realizes the real-time deployment of the deep learning algorithm of the domestically produced intelligent target detection platform， and explores the realization of an autonomous and controllable intelligent spaceborne front-end system.

Key words: on-board platform, remote sensing image, deep learning, domestic device, model optimization, ship detection