基于多尺度字符特征提取的场景文本识别算法研究  

Research on Scene Text Recognition Algorithm Based on Multi-scale Character Feature Extraction

作　　者：崔瑞雪 舒琪 王旭智[1,2] 万旺根[1,2] 孙学涛 张振 

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]上海大学智慧城市研究院,上海200444 [3]上海交通大学医学院附属仁济医院宝山分院,上海200444

出　　处：《工业控制计算机》2025年第3期59-61,共3页Industrial Control Computer

摘　　要：在场景文字识别任务中,字符特征的提取是至关重要的一环,对于提高识别准确率具有举足轻重的作用。为了提高模型在处理文字弯曲、拥挤、变形、模糊等复杂情况时的性能,提出了一种基于多尺度字符特征提取的场景文本识别算法,以表现优异的SVTR模型为基准,改进了模型中的Mixing Block,使用更丰富的混合块进行特征提取,使得模型能够在不同尺度上对各字符组件进行有效关系建模,从而更全面地理解字符间的联系和字符与全局之间的信息。最后在BCTR等数据集上进行训练和测试,实验结果表明,该方法在BCTR数据集上准确提升了2.8%,验证了其有效性。In order to improve the performance of the model in dealing with complex situations such as text bending,crowding,deformation and ambiguity,this paper proposes a scene text recognition algorithm based on multi-scale character feature extraction.Taking the SVTR model as the basis,the Mixing Block in the model is improved and more abundant mixing blocks are used for feature extraction.The model can effectively model the relationship of each character component at different scales,so as to understand the relationship between characters and the information between characters and the whole world more comprehensively.Finally,this paper conducts training and testing on BCTR and other data sets,and the experimental results showed that the proposed method accurately improved by 2.8%on the BCTR data set.

关 键 词：多尺度字符特征提取 文本识别 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]