首页|基于轻量化神经网络的石窟壁画破损检测方法

基于轻量化神经网络的石窟壁画破损检测方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
针对石窟壁画脱落与破损检测过程中存在检测精度低、实时性差的问题,提出了基于轻量化神经网络和多重注意力机制的石窟壁画破损检测方法.首先,引入Ghost Conv完成轻量化特征提取,降低模型复杂度;其次,加入双重注意力机制增加特征提取的倾向性,加快模型收敛速度;最后,使用加权双向特征金字塔拼接方式高效融合特征信息,通过复合缩放完成预测.实验结果表明:改进后的算法网络层数减少了34.40％.参数量和浮点运算量分别降低了62.98％和68.77％,模型体积压缩了62.78％.检测精度高达64.7％,实时检测速度从63.60帧/s提升至97.56帧/s,提高了约53.39％.
Damage Detection Method for Grotto Murals Based on Lightweight Neural Network
To address the issues of low detection precision and poor real-time performance in the process of grotto mural detachment and damage detection,we propose a grotto mural damage detection method based on a lightweight neural network and multiple attention mechanisms.First,Ghost Conv is introduced to complete lightweight feature extraction and reduce model complexity.Sec-ond,we add a double attention mechanism to increase the tendency of feature extraction and accel-erate model convergence.Finally,we use a weighted bidirectional feature pyramid network to effi-ciently fuse feature information and complete prediction by composite scaling.The experimental re-sults show that the improved algorithm reduces the number of network layers by 34.40％.The number of parameters and floating point operations are reduced by 62.98％and 68.77％,respec-tively,and the model volume is compressed by 62.78％.The detection precision is 64.7％,and the real-time detection speed is improved from 63.60 frame/s to 97.56 frame/s,which is approxi-mately 53.39％.

deep learningneural networklightweight modelattention mechanismmurals damage detection

吴利刚、张梁

展开 >

山西大同大学机电工程学院,山西大同 037003

大连海事大学信息科学与技术学院,辽宁大连 116026

山西大同大学煤炭工程学院,山西大同 037003

深度学习 神经网络 轻量化模型 注意力机制 壁画破损检测

山西省哲学社会科学规划项目(2021)山西大同大学科研项目云冈专项(2020)山西省高等学校科技创新项目(2021)校级科研专项云冈学研究项目(2021)

2021YY1982020YGZX014202113912021YGZX27

2024

10.13976/j.cnki.xk.2023.2456

信息与控制
中国自动化学会 中国科学院沈阳自动化研究所

信息与控制

CSTPCD北大核心
影响因子：0.576
ISSN：1002-0411
年,卷(期)：2024.53(1)
  • 21