结合多粒度特征融合的自然场景文本检测方法  

Natural Scene Text Detection Algorithm Combining Multi-granularity Feature Fusion

作　　者：陈卓[1] 王国胤[1] 刘群[1] CHEN Zhuo;WANG Guo-yin;LIU Qun(Chongqing Key Laboratory of Computational Intelligence,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学计算智能重庆市重点实验室,重庆400065

出　　处：《计算机科学》2021年第12期243-248,共6页Computer Science

基　　金：国家自然科学重点基金项目(61936001)。

摘　　要：自然场景下的文本信息通常具有多样性和复杂性的特点。由于采用手工设计特征的方式,传统的自然场景文字检测方法缺乏鲁棒性,而已有的基于深度学习的文本检测方法在各层网络提取特征的过程中存在丢失重要特征信息的问题。文中从多粒度和认知学的角度,提出了一种结合多粒度特征融合的自然场景文本检测方法。该方法的主要贡献是通过对通用特征提取网络的不同粒度特征进行融合,并加入残差通道注意力机制,使得模型在充分学习图像中不同粒度特征信息的基础上,更加关注目标特征信息并抑制无用的信息,提升了模型的鲁棒性和准确率。实验结果表明,相比其他最新的方法,该方法在公开数据集上取得了85.3%的准确率和82.53%的F值,具有更好的性能。In natural scenes, text information usually has the characteristics of diversity and complexity.Due to the way of manua-lly designing features, traditional natural scene text detection methods lack robustness, and the existing text detection methods based on deep learning have the problem of losing important feature information in the process of extracting features in each layer of the network.This paper proposes a natural scene text detection method combined with multi-granularity feature fusion.The main contribution of this method is that by combining the features of different granularities in the general feature extraction network and adding the residual channel attention mechanism, the model can pay more attention to the target feature information and suppress useless information on the basis of fully learning the feature information of different granularities in the image, and this method improves the robustness and accuracy of the model.The experimental results show that, compared with other latest me-thods, the model has achieved 85.3% accuracy and 82.53% F-value on public datasets, and has better performance.

关 键 词：特征提取 多粒度信息 残差注意力 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]