Abstract: With the rapid development of deep learning, object detection algorithms based on deep learning have been widely used in many fields. However, with the continuous evolution of the algorithm, a series of challenges has gradually emerged: the increase in the complexity of the model leads to the rapid expansion of the number of parameters and the amount of computation, which in turn reduces the running speed and makes it difficult to meet the needs of high-real-time application scenarios. The high requirements of the algorithm on hardware performance limit its efficient deployment in resource-constrained environments such as mobile devices and edge computing, and narrow its application range. The significant rise in training costs, including the need for large computational resources and long training time, also prevents the rapid iteration of models. In order to deal with these challenges, the research of lightweight object detection algorithm comes into being. This paper aims to review the recent progress of lightweight object detection algorithms based on deep learning. In this paper, the task of object detection and its evaluation index are summarized, and then the development history and representative model of object detection algorithm are reviewed in detail. On this basis, the paper focuses on the lightweight technology of object detection algorithm, including lightweight network architecture design to reduce model computational complexity and spatial complexity, lightweight convolutional technology innovation to reduce the number of parameters and computation while maintaining model performance, deep learning model compression method to optimize model structure to reduce storage requirements. Efficient deployment and real-time inference of resource-constrained devices are realized. Finally, this paper summarizes the current research status of lightweight object detection algorithm, and prospects and thinks in the aspects of multi-field technology integration, hardware architecture optimization and edge device deployment.

Key words: deep learning, object detection algorithm, model lightweight

摘要： 随着深度学习的快速发展，基于深度学习的目标检测算法在多个领域得到广泛应用。然而，随着算法不断演进，一系列挑战也逐渐浮现：模型复杂度增加导致参数量和计算量急剧膨胀，进而使得运行速度下降，难以满足高实时性应用场景需求；算法对硬件性能高要求，限制其在移动设备和边缘计算等资源受限环境中高效部署，缩小其应用范围；训练成本显著上升，包括对大量计算资源和长时间训练需求，也阻碍模型快速迭代。为了应对这些挑战，目标检测算法轻量化研究应运而生。综述基于深度学习的目标检测算法轻量化研究最新进展。对目标检测任务及其评价指标进行概述，详细回顾目标检测算法发展历程及代表性模型。在此基础上，重点探讨目标检测算法的轻量化技术，包括：轻量级网络架构设计，降低模型计算复杂性和空间复杂度；轻量级卷积技术创新，在减少参数量和计算量的同时保持模型性能；深度学习模型压缩方法，优化模型结构降低存储需求。实现资源受限设备高效部署与实时推理。总结当前轻量化目标检测算法研究现状，并在多领域技术融合、硬件架构优化以及边缘设备部署等方面进行展望与思考。

关键词: 深度学习, 目标检测算法, 模型轻量化