南京信息工程大学学报2026，Vol.18Issue(2)：221-230,10.DOI:10.13878/j.cnki.jnuist.20250221002
基于优化DBNet-CRNN的端子标识检测识别算法研究
An optimized DBNet-CRNN based algorithm for terminal detection and recognition
摘要
Abstract
To improve substation inspection efficiency,this paper proposes a deep learning-based method for de-tecting and recognizing distorted terminal block labels in substation.The detection model is built upon the DBNet framework,wherein the original ResNet backbone is replaced with ConvNeXt V2.This upgrade uses ConvNeXt V2's modern architectural design to significantly enhance the model's global information modeling and feature extraction capabilities for terminal identification.To further improve the detection accuracy for twisted and deformed terminal labels,we integrate an Efficient Multi-scale Attention(EMA)module and a Deformable Convolutional Network(DCNv4)to effectively strengthen the model's ability to capture global contextual information and enhance its ro-bustness to irregular text shapes.After optimization,the terminal detection model achieves F1 score of 97.4%,which represents an 18.9 percentage-point improvement over the original model,with only a 2.9%increase in calculation amount.For the recognition model,we take the Convolutional Recurrent Neural Network(CRNN)framework and introduce Spatial and Channel Convolution(SCConv)to optimize the feature extraction process,which significantly reduces redundant features and alleviates the computational burden.In the sequence modeling part,the selective state space network Mamba is used to replace the original Long Short-Term Memory(LSTM).Mamba's state space model and selective mechanism enable dynamic sequence data modeling,allowing it to adaptively focus on critical parts of the sequence and significantly enhance the capture of long sequence dependencies.The optimized recogni-tion model achieves an accuracy of 98.2%,which is 2.7 percentage points higher than the original model.Experi-mental results show that the proposed method exhibits excellent detection capability and recognition accuracy for substation terminal identification under challenging conditions such as distortion,weak light and blur.关键词
变电所巡检/文本检测/文本识别/端子标识/DBNet/卷积循环神经网络Key words
substation inspection/text detection/text recognition/terminal identification/DBNet/convolutional re-current neural network(CRNN)分类
信息技术与安全科学引用本文复制引用
王景琦,陈煜琦,薛强,赵瑞清,王硕禾..基于优化DBNet-CRNN的端子标识检测识别算法研究[J].南京信息工程大学学报,2026,18(2):221-230,10.基金项目
国家自然科学基金(12072203) （12072203）
国能朔黄铁路发展有限责任公司科研课题(SHTL-24-32) （SHTL-24-32）
