刘乐,张晓松,黄锋,方一鸣.基于改进DBNet和SVTR算法的连铸板坯号检测与识别[J].电子测量与仪器学报,2024,38(2):67-75 |
基于改进DBNet和SVTR算法的连铸板坯号检测与识别 |
Detection and recognition of continuous casting slab numbers based onimproved DBNet and SVTR algorithms |
|
DOI: |
中文关键词: 板坯号识别 DBNet 特征金字塔融合 端到端网络 SPIN矫正 SVTR |
英文关键词:slab number recognition DBNet feature pyramid fusion end-to-end network SPIN correction SVTR |
基金项目:秦皇岛市科学技术研究与发展计划项目(202302B048)、河北省创新能力提升计划项目(22567619H)、河北省教育厅科学研究项目(ZD2021102)资助 |
|
Author | Institution |
Liu Le | Key Laboratory of Intelligent Rehabilitation and Neromodulation of Hebei Province,
Yanshan University, Qinhuangdao 066004,China |
Zhang Xiaosong | Key Laboratory of Intelligent Rehabilitation and Neromodulation of Hebei Province,
Yanshan University, Qinhuangdao 066004,China |
Huang Feng | Key Laboratory of Intelligent Rehabilitation and Neromodulation of Hebei Province,
Yanshan University, Qinhuangdao 066004,China |
Fang Yiming | Key Laboratory of Intelligent Rehabilitation and Neromodulation of Hebei Province,
Yanshan University, Qinhuangdao 066004,China |
|
摘要点击次数: 285 |
全文下载次数: 450 |
中文摘要: |
针对钢铁连铸产线板坯号识别字符区域小、光照变化复杂、板坯号图像质量差等问题,提出了一种基于深度学习的连铸板坯号检测与识别两阶段算法。首先,基于采集的连铸产线板坯图像,制备用于板坯号检测与识别的数据集;其次,在板坯号检测阶段,基于DBNet算法设计一种AD-PAN特征融合结构,以增强检测算法的多尺度特征融合能力和扩大感受野,提高板坯号定位精度;再次,在板坯号识别阶段,引入SPIN矫正网络和SVTR板坯号识别网络进行端到端训练,使其能够主动转换输入亮度,并改善字符间以及字符与背景间色彩失真的问题。最后,在自制的板坯号检测与识别数据集上进行了对比实验。实验结果表明,本研究提出的算法能够有效定位辊道上不同位置的板坯,并且在复杂背景下对板坯号进行鲁棒识别。其中,板坯号检测Hmean数值为97.92%,板坯号识别的准确率为97.33%,验证了本文所提算法具有较高的板坯号检测与识别精度。 |
英文摘要: |
In order to tackle the challenges associated with small character regions, complex lighting variations, and poor image quality in the identification of slab numbers in steel continuous casting production lines, a two-stage algorithm is proposed for slab number detection and recognition, utilizing deep learning techniques. Firstly, datasets for slab number detection and recognition are prepared based on the collected slab images of a continuous casting production line. Secondly, in the slab number detection stage, an AD-PAN feature fusion structure based on the DBNet algorithm is designed to enhance the multi-scale feature fusion capability and expand the receptive field of the detection algorithm, thereby improving the localization accuracies of the slab numbers. Thirdly, in the slab number recognition stage, the proposed algorithm incorporates a SPIN correction network and a SVTR slab number recognition network for end-to-end training, enabling them to actively transform input brightness and improve color distortion among characters and between characters and backgrounds. Finally, comparative experiments are conducted on self-made datasets for slab number detection and recognition, and experimental results demonstrate that the algorithm proposed in this study is capable of effectively locating slab at different positions on the roller table and robustly recognizing slab numbers in complex backgrounds. The Hmean value for slab number detection is 97.92%, and the accuracy for slab number recognition is 97.33%, confirming the high accuracy of slab number detection and recognition achieved by the proposed algorithm. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|