吉林大学学报(工学版) ›› 2024, Vol. 54 ›› Issue (7): 2072-2079.doi: 10.13229/j.cnki.jdxbgxb.20230150

• 计算机科学与技术 • 上一篇    

一种工件表面压印字符识别网络

游新冬(),郭磊,韩晶(),吕学强   

  1. 北京信息科技大学 网络文化与数字传播北京市重点实验室,北京 100101
  • 收稿日期:2023-02-21 出版日期:2024-07-01 发布日期:2024-08-05
  • 通讯作者: 韩晶 E-mail:youxindong7895@126.com;hanjing@bistu.edu.cn
  • 作者简介:游新冬(1979-),女,副教授,博士.研究方向:计算机视觉.E-mail: youxindong7895@126.com
  • 基金资助:
    国家自然科学基金项目(62171043);北京市自然科学基金项目(4212020)

An character recognition network for imprint character

Xin-dong YOU(),Lei GUO,Jing HAN(),Xue-qiang LYU   

  1. Beijing Key Laboratory of Internet Culture and Digital Communication,Beijing Information Science and Technology University,Beijing 100101,China
  • Received:2023-02-21 Online:2024-07-01 Published:2024-08-05
  • Contact: Jing HAN E-mail:youxindong7895@126.com;hanjing@bistu.edu.cn

摘要:

工件表面的压印字符存在凹凸不平、锈蚀、风化等问题,导致传统的字符识别算法难以取得满意的效果。针对这一问题,将工件表面压印字符的识别视为一类特殊的目标检测问题,并针对其特性设计了一种两阶段识别网络:定位-分类网络。定位网络使用无锚框的方法提取字符感兴趣区域,有效解决了字符区域提取困难的问题。分类网络采用特征解耦的卷积模块和结构重参数化技术,能够在不增加额外参数的情况下提升分类的准确率。此外,分类网络采用跨域迁移学习的训练策略,能够有效解决实际应用中的小样本和类别不平衡问题。在自建螺栓数据集和SynthText数据集上的实验结果表明,该算法的整体精度能够达到98%和92%,优于对比算法。

关键词: 压印字符, 字符识别, 无锚框, 小样本, 目标检测

Abstract:

The imprint characters on the surface of the workpiece are uneven, rusty, and weathered, which the traditional character recognition methods hard to achieve satisfactory results. This paper regards the characters recognition task as a particular detection problem and designs a two-stage recognition network according to its characteristics: location and classification network. The location newtork uses the anchor-free method to extract the region of interest of characters, which effectively solves the problem of character region extraction. The classification network uses the Feature Decoupled Convolution Block and the Structural Re-parameterization technology, which can significantly improve the classification accuracy without any extra parameter. The transferring learning is used to solve the small sample problem and imbalance problem in the training stage. The experimental results on the self-built bolt dataset and the SynthText dataset show that the algorithm can achieve overall accuracies of 98% and 92%, respectively, which is superior to the compared algorithms.

Key words: imprint character, character recognition, anchor-free, small sample, object detection

中图分类号: 

  • TP391

图1

定位网络结构"

图2

标签生成"

图3

采用跨域迁移学习训练策略的分类网络"

图4

特征解耦的卷积模块"

图5

跨域迁移学习策略"

图6

定位和分类数据集"

表1

字符定位指标"

数据集算法准确率召回率F1
自建螺栓数据集Faster R-CNN0.9750.9620.969
YOLOv50.9790.9700.974
CenterNet0.9800.9490.964
RetinaNet0.9430.9060.924
本文0.9960.9970.996
SynthText数据集Faster R-CNN0.9550.9370.946
YOLOv50.9620.9390.950
CenterNet0.9450.9200.932
RetinaNet0.8980.8860.891
本文0.9730.9630.968

表2

FD Block对比实验结果"

数据集算 法基线准确率替换后准确率基线帧率/(帧·s-1替换后帧率/(帧·s-1
自建螺栓数据集VGG160.9630.974112.7113.9
ResNet500.9750.98195.195.8
SynthText数据集VGG160.9150.922184.8186.3
ResNet500.9230.932148.5149.1

表3

跨域迁移的提升"

跨域迁移
自建螺栓数据集0.9810.990
SynthText数据集0.9320.949

表4

不同算法实验对比"

算法准确率召回率F1
Faster R-CNN0.9470.9340.941
YOLOv50.9670.9560.962
CenterNet0.9610.9310.946
RetinaNet0.9230.8870.905
本文0.9850.9860.985

表5

不同算法实验对比"

算法准确率召回率F1
Faster R-CNN0.9040.8870.895
YOLOv50.9150.8930.903
CenterNet0.8770.8640.870
RetinaNet0.8430.8270.835
本文0.9230.9140.918

图7

效果展示"

1 黄慧宁, 张学军, 黄菊, 等. 基于深度学习YOLOv2算法的钢材压印字符识别研究[J]. 计算机科学与应用, 2020, 10(1): 126-135.
Huang Hui-ning, Zhang Xue-jun, Huang Ju. Research on steel stamping character recognition on deep learning YOLOv2 algorithm[J]. Computer Science and Application, 2020, 10(1): 126-135.
2 Chen X, Jin L, Zhu Y, et al. Text recognition in the wild: a survey[J]. ACM Computing Surveys(CSUR), 2021, 54(2): 1-35.
3 Ding X, Guo Y, Ding G, et al. ACNet: strengthening the kernel skeletons for powerful CNN via asymmetric convolution blocks[C]∥IEEE/CVF International Conference on Computer Vision(ICCV), Seoul, Korea(South), 2019: 1911-1920.
4 Chen Y, Liu S, Shen X, et al. Fast point R-CNN[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision,Seoul, Korea (South),2019: 9774-9783.
5 Qiao L, Zhao Y, Li Z, et al. Defrcn: decoupled faster r-cnn for few-shot object detection[C]∥Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, Canada, 2021: 8681-8690.
6 Oksuz K, Cam B C, Kalkan S, et al. Imbalance problems in object detection: a review[J]. IEEE Trans Pattern Anal Mach Intell, 2021(10): 3388-3415.
7 Cheng T, Wang X, Huang L, et al. Boundary-preserving mask R-CNN[C]∥European Conference on Computer Vision, Berlin: Springer, 2020: 660-676.
8 He K, Girshick R, Dollár P. Rethinking imagenet pre-training[DB/OL].[2023-01-26]..
9 Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[DB/OL].[2023-01-26]..
10 Ding X, Zhang X, Ma N, et al. RepVGG: making VGG-style convnets great again[C]//IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, USA,2021:No.01352.
11 Zhou Z, Siddiquee M, Tajbakhsh N, et al. UNet++: redesigning skip connections to exploit multiscale features in image segmentation[J]. IEEE Transactions on Medical Imaging, 2020, 39(6): 1856-1867.
12 Gupta A, Vedaldi A, Zisserman A. Synthetic data for text localisation in natural images[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 2315-2324.
13 Redmon J, Farhadi A. YOLOv3: an incremental improvement[DB/OL].[2023-01-27]..
14 Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[DB/OL].[2023-01-27]..
15 Duan K, Bai S, Xie L, et al. Centernet: keypoint triplets for object detection[DB/OL].[2023-01-27]..
16 He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 770-778.
[1] 高云龙,任明,吴川,高文. 基于注意力机制改进的无锚框舰船检测模型[J]. 吉林大学学报(工学版), 2024, 54(5): 1407-1416.
[2] 陈仁祥,胡超超,胡小林,杨黎霞,张军,何家乐. 基于改进YOLOv5的驾驶员分心驾驶检测[J]. 吉林大学学报(工学版), 2024, 54(4): 959-968.
[3] 张云佐,郭威,李文博. 遥感图像密集小目标全方位精准检测算法[J]. 吉林大学学报(工学版), 2024, 54(4): 1105-1113.
[4] 王宏志,宋明轩,程超,解东旋. 基于改进YOLOv4-tiny算法的车距预警方法[J]. 吉林大学学报(工学版), 2024, 54(3): 741-748.
[5] 李晓旭,安文娟,武继杰,李真,张珂,马占宇. 通道注意力双线性度量网络[J]. 吉林大学学报(工学版), 2024, 54(2): 524-532.
[6] 王春华,李恩泽,肖敏. 多特征融合和孪生注意力网络的高分辨率遥感图像目标检测[J]. 吉林大学学报(工学版), 2024, 54(1): 240-250.
[7] 薛珊,张亚亮,吕琼莹,曹国华. 复杂背景下的反无人机系统目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 891-901.
[8] 陶博,颜伏伍,尹智帅,武冬梅. 基于高精度地图增强的三维目标检测算法[J]. 吉林大学学报(工学版), 2023, 53(3): 802-809.
[9] 刘晶红,邓安平,陈琪琪,彭佳琦,左羽佳. 基于多重注意力机制的无锚框目标跟踪算法[J]. 吉林大学学报(工学版), 2023, 53(12): 3518-3528.
[10] 黄彭奇子,段晓君,黄文伟,晏良. 基于元学习的小样本图像非对称缺陷检测方法[J]. 吉林大学学报(工学版), 2023, 53(1): 234-240.
[11] 高明华,杨璨. 基于改进卷积神经网络的交通目标检测方法[J]. 吉林大学学报(工学版), 2022, 52(6): 1353-1361.
[12] 曲优,李文辉. 基于锚框变换的单阶段旋转目标检测方法[J]. 吉林大学学报(工学版), 2022, 52(1): 162-173.
[13] 曹洁,屈雪,李晓旭. 基于滑动特征向量的小样本图像分类方法[J]. 吉林大学学报(工学版), 2021, 51(5): 1785-1791.
[14] 潘德伦,冀隽,张跃进. 基于运动矢量空间编码的视频监控动态目标检测方法[J]. 吉林大学学报(工学版), 2021, 51(4): 1370-1374.
[15] 金立生,郭柏苍,王芳荣,石健. 基于改进YOLOv3的车辆前方动态多目标检测算法[J]. 吉林大学学报(工学版), 2021, 51(4): 1427-1436.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!