基于卷积神经网络的文本检测算法研究  

The Research of Text Detection Based on Convolutional Neural Network

作　　者：李阳 李绍彬[1] 解云超 冯爽 LI Yang;LI Shao-bin;XIE Yun-chao;FENG Shuang(School of Information Engineering,Communication University of China, Beijing 100024 ,China)

机构地区：[1]中国传媒大学信息工程学院,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2019年第1期70-76,共7页Journal of Communication University of China：Science and Technology

摘　　要：随着深度学习的发展,利用神经网络对文本进行检测得到了更深入的研究和更广泛的应用。本文基于Text-Boxes算法,在考虑到足球赛事场景下的文本特点后对其进行改进,提升了该场景下文本检测的效果。针对足球赛事场景下的文本几何形状多样性的特点,设置适应于足球场景中文本检测的默认框;针对影响模型优化的样本不均衡问题,使用Focal Loss作为用于分类的损失函数;最后使用非极大值抑制过滤冗余的矩形框,获得最终的检测结果。本文自行标注了足球赛事场景数据集,用于网络的训练,验证了本文算法的有效性。With the development of deep learning,text detection based on neural network has gained more in-depth research and more extensive application. Considering the characteristics of the text in the football match scene,a novel neural network architecture is proposed based on the TextBoxes. The proposed algorithm performs well in the task of text detection in the football match scene. According to the text geometry diversity in the football match scene,we select proper default boxes which fit the text boxes in football match scene. In order to solve the sample imbalance problem that interference model optimization,we propose Focal Loss as a loss function for classification. Finally,the Non-maximal suppression is applied to eliminate redundant bounding boxes and aggregate outputs. We make a dataset for training,and verify the effectiveness of the algorithm.

关 键 词：文本检测 卷积神经网络 数据集 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]