基于深度网络模型压缩与加速的高分辨率遥感图像在轨目标检测

Intelligent on-orbit object detection in high-resolution remote sensing images is crucial for military and civilian fields such as reconnaissance and emergency rescue. With powerful feature learning and representation ability, object detection methods based on deep learning can significantly improve the detection accuracy. However, deep models are always with large parameters, high storage and computational cost, and are difficult to be deployed in on-orbit platform with limited storage, processing and energy consumption capacity. To this end, taking the parameter redundancy and parallelism characteristics of deep models into account, this project proposes an on-orbit object detection method in high-resolution remote sensing images based on deep model compression and acceleration, with the principle of “reducing the number of parameters, simplified parameter representation and parallel computing acceleration”. The method reduces redundant parameters through filter pruning considering energy consumption and the bit numbers of parameters through layer-by-layer non-uniform quantization and encoding, followed by a parallel computing implementation deployed on embedded heterogeneous systems. The research of this project is of great significance for maximizing the effective data transmission under limited satellite-ground station link, and reducing response latency of object detection in massive high-resolution remote sensing data.

高分辨率遥感图像在轨智能目标检测对于情报侦察预警、应急救援等军事和民用领域至关重要。基于深度学习的目标检测算法特征学习和表达能力强，能够显著提升检测精度，但其模型参数多、计算复杂度高，与在轨平台对存储、计算、能耗的限制形成巨大矛盾。因此，本项目基于深度网络模型的参数冗余性和算法可并行性，以“降低参数数量-精简参数表达-并行计算加速”为主线，提出基于深度网络模型压缩与加速的高分辨率遥感图像在轨目标检测方法：基于顾及能耗的滤波器剪枝，减少冗余参数；通过参数非均匀量化与编码，降低参数存储位宽；在以上算法层面降低模型大小和复杂度的基础上，形成嵌入式异构平台上的遥感图像目标检测并行计算方法，降低响应延迟。本项目的研究对于实现有限带宽下传输有效数据最大化、提高海量高分辨率遥感数据目标检测的时效性具有重要意义。

高分辨率遥感图像在轨智能目标检测对于情报侦察预警、应急救援等军事和民用领域至关重要。近年来，基于深度学习的目标检测算法得到长足发展，其特征学习和表达能力强，能够显著提升检测精度。然而，这些现有方法是典型的计算密集和访存密集型过程，涉及的参数量大，存储、计算和能耗代价高，难以部署和运行于存储、计算、能耗等受限的在轨平台。因此，本项目探索高分辨率遥感图像目标检测深度网络模型的参数冗余性和算法可并行性，以“降低参数数量-精简参数表达-并行计算加速”为研究主线，发展了一种基于深度网络模型压缩与加速的高分辨率遥感图像在轨目标检测方法框架，具体包括：发展了一种融合注意力机制的高分辨率遥感图像旋转目标检测算法、一种基于紧凑卷积模块和滤波器剪枝的轻量化遥感图像目标检测算法、一种能耗引导的逐层剪枝率估计和低能耗遥感图像目标检测算法、一种基于资源受限嵌入式异构平台的轻量化遥感图像目标检测方法、一种基于多节点分布式集群的遥感图像目标检测并行与优化策略，构建了一个遥感图像快速目标检测原型系统。本项目研究能够大幅降低遥感目标检测深度网络模型的冗余参数、模型大小、计算复杂度、响应延迟、推理能耗，从而满足存储、计算、能耗等资源受限的在轨平台智能目标检测需求。项目研究成果对于实现有限带宽下传输有效数据最大化，提高海量高分辨率遥感数据目标检测的时效性，进一步拓展至灾害实时监测与应急管理、无人机精准农业监测等领域具有重要意义。