首页|基于改进EAST算法的电气设备铭牌文字检测

基于改进EAST算法的电气设备铭牌文字检测

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
针对电气设备铭牌文本检测任务中存在检测效果较差、文本边界框检测不准确等问题,提出一种基于深度学习算法EAST的改进算法,重点优化EAST算法对长文本检测效果较差和小目标遗漏问题.使用ResNet50残差网络代替VGG算法作为EAST算法特征提取的主干网络,同时引入金字塔注意力模块和空洞卷积,增强图像各层特征图的细节信息和彼此间的关联性、扩大特征图的感受野,解决EAST算法存在的缺陷问题.此外,使用Dice、Focal损失建立改进算法的损失函数,对改进模型的训练过程进行优化.结果表明,改进的算法在ICDAR 2015以及自行建立的电气设备铭牌数据集上的召回率和准确率均高于传统EAST算法,召回率平均提升约7.8%,准确率平均提升4.3%,综合性能提升约6.3%.
Text Detection of Electrical Equipment Nameplate Based on Improved EAST Algorithm
In response to the problems of poor detection performance and inaccurate text bounding box detection in the text detection task of electrical equipment nameplates,an improved algorithm based on the deep learning algorithm EAST has been proposed,focusing on optimizing the EAST algorithm's poor detection performance for long texts and small target omissions.ResNet50 residual network is used to replace VGG algorithm as the backbone network for feature extraction of EAST algorithm.At the same time,pyramid attention module and void convolution are introduced to enhance the detail information and correlation between the feature maps of each layer of the image,and expand the receptive field of the feature map,and address the defects of EAST algorithm.In addition,Dice and Focal losses are used to establish the loss function of the improved algorithm to optimize the training process of the improved model.The results show that the improved algorithm has higher recall and accuracy than the traditional EAST algorithm on ICDAR 2015 and the self-established electrical equipment nameplate dataset,with an average improvement of about 7.8%in recall and 4.3%in accuracy,and an overall performance improvement of about 6.3%.

nameplate of electrical equipmenttext detectionEASTfeature extractionloss function

刘彦希、吴浩、蔡源、唐丹、宋弘

展开 >

四川轻化工大学自动化与信息工程学院,四川 宜宾 644000

人工智能四川省重点实验室,四川 宜宾 644000

阿坝师范学院,四川 阿坝 623002

电气设备铭牌 文本检测 EAST 特征提取 损失函数

四川省科技厅项目四川省科技厅项目四川省科技厅项目自贡市科技局项目人工智能四川省重点实验室项目四川轻化工大学人才引进项目四川轻化工大学研究生创新基金项目

2021YFG03132022YFS05182022ZHCG00352020YGJC162020RZY032021RC12Y2022122

2024

10.11863/j.suse.2024.03.06

四川轻化工大学学报(自然科学版)
四川理工学院

四川轻化工大学学报(自然科学版)

影响因子：0.44
ISSN：2096-7543
年,卷(期)：2024.37(3)