基于轻量级网络的藏文复杂背景文字识别方法研究

doi:10.11996/JG.j.2095-302X.2026030492

图学学报 ›› 2026, Vol. 47 ›› Issue (3): 492-499.DOI: 10.11996/JG.j.2095-302X.2026030492

• 图像处理与计算机视觉 • 上一篇下一篇

基于轻量级网络的藏文复杂背景文字识别方法研究

王悦凝¹^,², 才让当知¹^,³(), 昝嵘¹, 孟磊¹

¹ 青海民族大学智能科学与工程学院，青海西宁 810007
² 青海民族大学公共计算机实验教学中心，青海西宁 810007
³ 藏语智能全国重点实验室，青海西宁 810008

收稿日期:2025-04-29 接受日期:2025-10-29 出版日期:2026-06-30 发布日期:2026-06-30
通讯作者:才让当知，E-mail：2056617785@qq.com
基金资助:
省部共建藏语智能信息处理及应用国家重点实验室项目(2024-Z-002)

Research on Tibetan complex background text recognition method based on lightweight network

WANG Yuening¹^,², CAIRANG Dangzhi¹^,³(), ZAN Rong¹, MENG Lei¹

¹ School of Intelligent Science and Engineering, Qinghai Minzu University, Xining Qinghai 810007, China
² Public Computer Experimental Teaching Demonstration Center, Qinghai Minzu University, Xining Qinghai 810007, China
³ The StateThe State Key Laboratory of Tibetan Intelligence, Xining Qinghai 810008, China

Received:2025-04-29 Accepted:2025-10-29 Published:2026-06-30 Online:2026-06-30
Contact: CAIRANG Dangzhi，E-mail：2056617785@qq.com
Supported by:
State Key Laboratory Project of Tibetan Intelligent Information Processing and Application jointly built by the province and ministry(2024-Z-002)

摘要/Abstract

摘要：

复杂背景下的藏文文字识别是藏文OCR领域的重要课题之一。针对传统的深度学习模型不仅参数量过大且不易部署，从而影响藏文文字识别效率的问题，提出了基于轻量级MobileNetV3-Gobal的藏文复杂背景文字识别模型。将基线模型MobileNetV3的最后一层引入了全局注意力机制的全局上下文模块（GCB），在保持网络模型计算量的同时通过全局注意力机制对藏文复杂背景的行文本图像进行全局建模，更易特征提取。实验结果表明，改进后的轻量级 MobileNetV3 - Global 藏文复杂背景文字识别模型有较好的准确率，通过自然场景和流媒体2种数据集测试，其准确率分别高达97.40%和96.15%，同时模型大小仅为6.05 MB，具有较高的实用性与部署潜力。

关键词: 藏文文字识别, MobileNetV3, Global Context Block, 特征提取, 轻量化

Abstract:

The recognition of Tibetan characters in complex backgrounds is one of the important research topics in the field of Tibetan OCR (Optical character Recognition). However, traditional deep learning models often have excessive parameter counts and are difficult to deploy, affecting the efficiency of Tibetan character recognition. To address the above problems, a Tibetan character recognition model for complex backgrounds based on the lightweight MobileNetV3-Global was proposed. The global attention mechanism (Global Context Block, GCB) was introduced into the last layer of the baseline model MobileNetV3. While maintaining the computational complexity of the network model, the global attention mechanism was used to perform global modeling on the line-text images of Tibetan characters in complex backgrounds, thereby facilitating feature extraction. Experimental results demonstrated that the improved lightweight MobileNetV3-Global model achieved high accuracy for Tibetan character recognition in complex backgrounds. In this study, tests were conducted on two types of datasets, namely natural scenes and streaming media. The accuracy rates of the model on these datasets reached 97.40% and 96.15%, respectively, and the model size was only 6.05 MB, indicating high practicality and deployment potential.

Key words: Tibetan character recognition, MobileNetV3, global context block, feature extraction, lightweight

中图分类号:

王悦凝, 才让当知, 昝嵘, 孟磊. 基于轻量级网络的藏文复杂背景文字识别方法研究[J]. 图学学报, 2026, 47(3): 492-499.

WANG Yuening, CAIRANG Dangzhi, ZAN Rong, MENG Lei. Research on Tibetan complex background text recognition method based on lightweight network[J]. Journal of Graphics, 2026, 47(3): 492-499.

图/表 13

图1 改进的MobileNetV3-Global Context Block模型总体结构

Fig. 1 The overall structure of the improved MobileNetV3-Global Context Block model

图2 Residual Unit结构

Fig. 2 Structure of the Residual Unit

图3 Global Context Block结构

Fig. 3 Structure of the Global Context Block

图4 2类藏文文字识别数据集样本示例((a) 自然场景；(b) 流媒体)

Fig. 4 Examples of samples from two types of tibetan character recognition datasets ((a) Natural scene; (b) Streaming media)

图5 与其他方法比较模型大小

Fig. 5 Comparison of model sizes with other methods

表1 与其他方法的对比实验

Table 1 Comparative experiments with other methods

方法	识别对象	模型	准确率/%	模型大小/MB
文献[1]	自然场景文本	PFRE	80.23	—
文献[5]		CNN + Attention	88.49	16
文献[6]		CNN-Bi-LSTM	90.55	41.2
文献[12]		CRNN+CTC	93.24	—
文献[7]	流媒体文本	CRNN	78.90	—
本文	自然场景文本	MobileNetV3-Gobal	97.40	6.05
本文	流媒体文本	MobileNetV3-Gobal	96.15	6.05

图6 2种藏文文字识别数据集的文本图像尺寸分布((a) 固定高度的自然场景；(b) 高度不固定的流媒体)

Fig. 6 Distribution of text image dimensions in the Tibetan - Chinese character recognition dataset ((a) Natural scene text recognition with fixed height; (b) Streaming media text recognition with variable height)

图7 全局注意力机制在2种数据集上的消融实验((a) 自然场景数据集；(b) 流媒体数据集)

Fig. 7 Experimental ablation of global attention mechanism on two datasets ((a) Natural scene datasets; (b) Streaming media datasets)

图8 场景下藏文文字识别训练集的准确率变化

Fig. 8 Accuracy changes of the training set for Tibetan text recognition in natural scenes

图9 藏文自然场景文字识别测试集上的相似度得分升降趋势

Fig. 9 Normalized edit distance in Tibetan text recognition under natural scenes

图10 自然场景下藏文文字识别训练集的准确率变化

Fig. 10 Accuracy changes of the training set for Tibetan text recognition in natural scenes

图11 藏文流媒体文字识别测试集上的相似度得分升降趋势

Fig. 11 Normalized edit distance in Tibetan text recognition for streaming media

图12 2类藏文文字图片识别结果样本示例((a) 自然场景；(b) 流媒体)

Fig. 12 Example of recognition results for two types of Tibetan characters ((a) Natural scenes; (b) Streaming media)

参考文献 20

[1]	才让当知. 印刷藏文吾美体检测与识别研究[D]. 西宁: 青海师范大学, 2024.
	CAIRANG D Z. Research on detection and recognition of printed tibetan cursive script[D]. Xining: Qinghai Normal University, 2024(in Chinese).
[2]	陈洋. 安卓平台上印刷体藏文识别软件的设计与实现[D]. 兰州: 西北民族大学, 2020.
	CHEN Y. Design and implementation of printed tibetan recognition software on android platform[D]. Lanzhou: Northwest Minzu University, 2020(in Chinese).
[3]	陈明, 陈炳权, 夏蓉. 基于轻量级Transformer模型的秦简文字识别算法[J]. 计算机仿真, 2025, 42(4): 459-467.
	CHEN M, CHEN B Q, XIA R. Qin bamboo slips character recognition algorithm based on lightweight transformer model[J]. Computer Simulation, 2025, 42(4): 459-467(in Chinese).
[4]	丁广刘. 基于轻量级CNN的自然场景文本检测和识别算法研究及其实现[D]. 苏州: 苏州大学, 2022.
	DING G L. Research and implementation of scene text detection and recognition algorithm based on lightweight CNN[D]. Suzhou: Soochow University, 2022(in Chinese).
[5]	童攀, 龙炳鑫, 拥措. 基于注意力机制藏文乌金体古籍文字识别研究[J]. 计算机技术与发展, 2023, 33(10): 163-168, 208.
	TONG P, LONG B X, YONG C. Research on Tibetan Ujin ancient book character recognition based on attention mechanism[J]. Computer Technology and Development, 2023, 33(10): 163-168, 208 (in Chinese).
[6]	拥措, 龙炳鑫, 拉毛杰, 等. 面向社交媒体的藏文图文多字体检测与识别研究[J]. 高原科学研究, 2023, 7(4): 76-85.
	YONG C, LONG B X, LAMAO J, et al. Study on the multi-font detection and recognition of Tibetan in the social media[J]. Plateau Science Research, 2023, 7(4): 76-85(in Chinese).
[7]	索朗曲珍. 视频中藏文文本的检测与识别方法研究[D]. 拉萨: 西藏大学, 2024.
	SUOLANG Q Z. Research on detection and recognition methods for tibetan text in videos[D]. Lhasa: Tibet University, 2024(in Chinese).
[8]	YUE X Y, KUANG Z H, LIN C H, et al. RobustScanner: dynamically enhancing positional clues for robust text recognition[C]//The 16th European Conference on Computer Vision. Cham: Springer, 2020: 135-151.
[9]	WAN Z Y, HE M H, CHEN H R, et al. TextScanner: reading characters in order for robust scene text recognition[EB/OL]. [2025-01-02]. https://ojs.aaai.org/index.php/AAAI/article/view/6891.
[10]	仁青东主, 尼玛扎西. 基于深度学习的自然场景藏文识别研究[J]. 高原科学研究, 2019, 3(4): 96-103.
	RENQING D Z, NIMA Z X. Study on Tibetan text recognition in natural scenes based on deep learning[J]. Plateau Science Research, 2019, 3(4): 96-103(in Chinese).
[11]	FU Z L, XIE H T, JIN G Q, et al. Look back again: dual parallel attention network for accurate and robust scene text recognition[C]//2021 International Conference on Multimedia Retrieval. New York: ACM, 2021: 638-644.
[12]	FANG S C, XIE H T, WANG Y X, et al. Read like humans: autonomous, bidirectional and iterative language modeling for scene text recognition[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 7094-7103.
[13]	侯闫. 多字体印刷藏文的文本检测与识别研究[D]. 拉萨: 西藏大学, 2023.
	HOU Y. Research on text detection and recognition of multi fonts Printed Tibetan[D]. Lhasa: Tibet University, 2023(in Chinese).
[14]	徐诗康, 刘俊峰, 曾君, 等. 基于跨模态和循环分解自注意力的场景文本识别[J]. 计算机工程与应用, 2025, 61(11): 176-184. DOI
	XU S K, LIU J F, ZENG J, et al. Scene text spotting based on cross-modal and circular factorized self-attention[J]. Computer Engineering and Applications, 2025, 61(11): 176-184(in Chinese). DOI
[15]	LI M H, LV T C, CHEN J Y, et al. TrOCR: transformer-based optical character recognition with pre-trained models[EB/OL]. [2025-01-28]. https://ojs.aaai.org/index.php/AAAI/article/view/26538.
[16]	YAN R J, PENG L R, XIAO S Y, et al. Primitive representation learning for scene text recognition[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 284-293.
[17]	YOUSEF M, HUSSAIN K F, MOHAMMED U S. Accurate, data-efficient, unconstrained text recognition with convolutional neural networks[J]. Pattern Recognition, 2020, 108: 107482. DOI URL
[18]	DU Y K, CHEN Z N, JIA C Y, et al. SVTR: scene text recognition with a single visual model[EB/OL]. [2025-01-02]. https://www.ijcai.org/proceedings/2022/124.
[19]	LITMAN R, ANSCHEL O, TSIPER S, et al. SCATTER: selective context attentional scene text recognizer[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 11959-11969.
[20]	HUANG M X, LIU Y L, PENG Z H, et al. SwinTextSpotter: scene text spotting via better synergy between text detection and text recognition[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 4583-4593.

基于轻量级网络的藏文复杂背景文字识别方法研究

Research on Tibetan complex background text recognition method based on lightweight network

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 13

参考文献 20

相关文章 15

编辑推荐

Metrics

本文评价

[1]	张宙, 王泽宇, 宋海玉, 李威, 葛鸣宇, 王嘉宇, 王文琦. 基于三维流形拟合与分频引导注意力机制的多聚焦图像融合[J]. 图学学报, 2026, 47(2): 351-359.
[2]	李晔, 贾俊洋, 黄冠, 李玉洁, 齐文婷, 刘岩. 用于夜视辅助驾驶的轻量化图像眩光去除方法[J]. 图学学报, 2026, 47(1): 57-67.
[3]	杨彪, 王学, 官铮, 龙萍. BSD-YOLO：基于动态稀疏注意力与自适应检测头的小目标车辆检测方法[J]. 图学学报, 2026, 47(1): 99-110.
[4]	贺蒙蒙, 张小艳, 李洪安. 基于Mamba结构的轻量级皮肤病变图像分割网络[J]. 图学学报, 2025, 46(6): 1257-1266.
[5]	杨杰, 李琮, 胡庆浩, 陈显达, 王云鹏, 刘晓晶. 面向轻量卷积神经网络的训练后量化方法[J]. 图学学报, 2025, 46(4): 709-718.
[6]	郭业才, 胡晓伟, 毛湘南. 多尺度密集交互注意力残差真实图像去噪网络[J]. 图学学报, 2025, 46(2): 279-287.
[7]	王志东, 陈晨阳, 刘晓明. 基于轻量化改进YOLOv8的通信光缆缺陷检测[J]. 图学学报, 2025, 46(1): 28-34.
[8]	李盛涛, 侯立群, 董亚松. 基于R-YOLOv7和MIMO-CTFNet的指针式仪表自动读数方法[J]. 图学学报, 2024, 45(6): 1313-1327.
[9]	闫建红, 冉同霄. 基于YOLOv8的轻量化无人机图像目标检测算法[J]. 图学学报, 2024, 45(6): 1328-1337.
[10]	翟永杰, 李佳蔚, 陈年昊, 王乾铭, 王新颖. 融合改进Transformer的车辆部件检测方法[J]. 图学学报, 2024, 45(5): 930-940.
[11]	曾志超, 徐玥, 王景玉, 叶元龙, 黄志开, 王欢. 基于SOE-YOLO轻量化的水面目标检测算法[J]. 图学学报, 2024, 45(4): 736-744.
[12]	朱强军, 胡斌, 汪慧兰, 王杨. 基于轻量化YOLOv8s交通标志的检测[J]. 图学学报, 2024, 45(3): 422-432.
[13]	郭宗洋, 刘立东, 蒋东华, 刘子翔, 朱熟康, 陈京华. 基于语义引导神经网络的人体动作识别算法[J]. 图学学报, 2024, 45(1): 26-34.
[14]	苑朝, 赵亚冬, 张耀, 王嘉璇, 徐大伟, 翟永杰, 朱松松. 基于YOLO轻量化的多模态行人检测算法[J]. 图学学报, 2024, 45(1): 35-46.
[15]	王欣雨, 刘慧, 朱积成, 盛玉瑞, 张彩明. 基于高低频特征分解的深度多模态医学图像融合网络[J]. 图学学报, 2024, 45(1): 65-77.