北京大学学报（自然科学版） ›› 2025, Vol. 61 ›› Issue (5): 884-890.DOI: 10.13209/j.0479-8023.2025.077

上一篇    下一篇

一种基于双级剪枝和训练后量化的火灾检测轻量化模型设计方法

徐鹏涛1,†, 王刚1, 张连杰1, 王越2, 黄华1   

  1. 1. 西安中核核仪器股份有限公司, 西安 710061 2. 北京大学软件与微电子学院, 北京 102600
  • 收稿日期:2024-08-04 修回日期:2024-10-21 出版日期:2025-09-20 发布日期:2025-09-20
  • 通讯作者: 徐鹏涛, E-mail: xupengtao(at)pku.edu.cn

A Lightweight Model Design Method for Fire Detection via Dual Level Pruning and Post-Training Quantization

XU Pengtao1,†, WANG Gang1, ZHANG Lianjie1, WANG Yue2, HUANG Hua1   

  1. 1. Xi’an Nuclear Instrument Co Ltd, Xi’an 710061 2. School of Software and Microelectronics, Peking University, Beijing 102600
  • Received:2024-08-04 Revised:2024-10-21 Online:2025-09-20 Published:2025-09-20
  • Contact: XU Pengtao, E-mail: xupengtao(at)pku.edu.cn

PDF

摘要：

针对火灾检测领域对高效、轻量级模型的迫切需求, 以SSD目标检测算法为基础, 搭建一种火灾检测轻量化模型。为减小模型规模, 提高计算速度以及满足实际场景下的部署要求, 采用剪枝和量化两种方法实现检测模型的轻量化。为实现模型网络在通道级和层级同时进行有效剪裁, 提出一种基于可融合残差卷积块的双级剪枝方法。为了有效地提升该轻量化模型的性能, 引入自适应方法, 实现一种基于自适应离群值去除的训练后量化方法。实验结果表明, 与原始方法相比, 所提剪枝方法和量化方法表现出明显的优势, 可在几乎不影响模型性能的情况下, 显著地减小模型规模, 同时保证火灾检测轻量化模型具有优异的性能。

关键词: 火灾检测, SSD目标检测, 模型轻量化, 剪枝, 量化

Abstract:

A lightweight fire detection model is designed to meet the urgent demand for efficient and lightweight models in the field of fire detection. The model is built based on the SSD object detection algorithm, and pruning and quantization methods are used to achieve lightweighting of the detection model in order to reduce model size, improve model speed, and meet the deployment requirements in practical scenarios. In order to achieve effective pruning of the model network at both channel and layer levels, a dual pruning method based on fusible residual convolution blocks is proposed. In order to effectively improve the performance of the quantization model, an adaptive method is introduced to quantize the model, which realizes a post-training quantization method based on adaptive outlier removal. The experimental results show that the proposed pruning method and quantization method exhibit significant advantages compared with the original method, and can significantly reduce the model size with almost no impact on performance. The final lightweight fire detection model also has excellent performance.

Key words: fire detection, SSD object detection, model lightweighting, pruning, quantization

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司