【摘要】 近年来，卷积神经网络（Convolutional Neural Network,CNN）、视觉Transformer以及大规模语言模型（Large Language Model,LLM）等深度学习模型在表示学习与任务泛化方面取得显著进展。然而，随着模型参数规模与计算复杂度迅速增加，其推理延迟与存储开销持续攀升。在遥感领域，高分辨率遥感图像与合成孔径雷达（Synthetic Aperture Radar,SAR）图像具有幅面大、分辨率高、场景复杂、目标多尺度以及小目标占比高等特点。因此，目标检测模型不仅需具备强大的表征能力以应对复杂背景与尺度变化，还必须满足星载、机载及边缘设备等受限平台对实时性与低功耗的严格要求，这使得“精度提升”与“高效部署”之间的矛盾日益凸显。本文首先系统回顾了深度网络压缩与加速技术的发展历程，涵盖从传统手工优化到剪枝、量化与知识蒸馏等经典方法，并进一步探讨了大模型时代下的条件计算与推理优化策略，包括混合专家系统（Mixture of Experts,MoE）、键值（Key Value,KV）缓存优化与混合量化等。其次，通过整合遥感目标检测领域的主流数据集与评价指标体系，本文构建了一个面向“精度-效率-资源”权衡的多维评估基准，用于分析不同压缩方案在遥感场景中的适配性与局限性。与现有综述多沿单一算法主线或仅聚焦轻量网络不同，本文从“算法视角”与“模型规模视角”两个互补维度组织研究脉络：算法视角侧重于在既定模型上通过剪枝、量化、蒸馏与稀疏化来降低计算与存储复杂度；模型规模视角则区分轻量模型与大规模基础模型在结构冗余、敏感模块与内存瓶颈方面的差异，进而分析压缩策略在不同体量模型中的作用机制与优化潜力。再次，基于这一双视角框架，本文对代表性方法的核心思想与实现要点进行对比，总结影响压缩效果的关键因素，并提炼可复用的工程化实践建议，以期为模型选型与部署配置提供参考。最后，结合遥感智能解译的应用需求与绿色低碳发展趋势，本文展望自适应压缩、多模态协同压缩、参数高效微调与软硬件协同优化等方向，为遥感目标检测模型在边缘端与工程系统中的高效落地提供系统化的理论支持与实践指导。更多还原

【Abstract】 In recent years, deep learning models such as convolutional neural networks(CNNs), vision transformers, and large language models(LLMs) have made remarkable breakthroughs in representation learning and task generalization. However, this progress has been accompanied by rapid increase in parameter scale and computational complexity, leading to rising inference latency and storage overhead. In remote sensing applications, high-resolution remote sensing images and synthetic aperture radar(SAR) images are characterized by large coverage, high resolution, complex scenes, and multi-scale variations with a high proportion of small objects. Object detection models must therefore possess strong representational capability to handle complex backgrounds and scale variations, while also meeting stringent real-time and low-power constraints on resource-limited platforms such as satellite-borne, airborne, and edge devices. As a result, the tension between “improving accuracy” and “efficient deployment” has become increasingly pronounced. This study systematically reviews the development of deep network compression and acceleration techniques, covering traditional manual optimizations as well as classical methods such as pruning, quantization, and knowledge distillation. It further discusses conditional computation and inference optimization in the era of large models, including mixture-of-experts(MoE) architectures, key-value(KV) cache optimization, and hybrid quantization. By integrating mainstream datasets and evaluation metrics in the field of remote sensing object detection, this paper establishes a multi-dimensional evaluation benchmark oriented toward the “accuracy-efficiency-resource” trade-offs, aimed at characterizing the applicability and limitations of different compression approaches in remote sensing scenarios. Unlike existing surveys that often follow a single algorithmic thread or focus solely on lightweight networks, this review organizes the research landscape from two complementary perspectives: the algorithmic perspective and the model-scale perspective. The former emphasizes reducing computational and storage complexity through pruning, quantization, distillation, and sparsification on a given model. The latter distinguishes between lightweight models and large-scale foundation in terms of structural redundancy, sensitive modules, and memory bottlenecks, analyzing how compression strategies operate and where further optimization is possible across models of different sizes. Based on this dual-perspective framework, the paper compares and analyzes the core ideas and implementation details of representative methods, summarizes key factors influencing compression performance, and extracts reusable engineering practices to inform model selection and deployment configurations. Finally, considering the application demands of remote sensing intelligent interpretation and the trend toward green and low-carbon computing, this paper looks ahead to promising directions such as adaptive compression, multimodal collaborative-compression, parameter-efficient fine-tuning, and hardware-software co-optimization. The aim is to provide systematic reference and theoretical support for the efficient deployment of remote sensing object detection models on edge devices and in engineering systems.更多还原