基于改进TextSnake的印章字符检测算法  

Stamp character detection algorithm based on improved TextSnake

作　　者：甘辉鑫 巩荣芬[1] 储茂祥[1] 杨永辉[1] GAN Huixin;GONG Rongfen;CHU Maoxiang;YANG Yonghui(School of Electronic and Information Engineering,University of Science and Technology Liaoning,Anshan 114051,China)

机构地区：[1]辽宁科技大学电子与信息工程学院,辽宁鞍山114051

出　　处：《辽宁科技大学学报》2024年第3期213-220,共8页Journal of University of Science and Technology Liaoning

基　　金：国家自然科学基金资助项目(71771112);辽宁省自然科学基金资助项目(2022-MS-353);辽宁省教育厅项目(2020LNZD06)。

摘　　要：针对字符模糊、宽高比多变、与背景文字重叠等因素导致的印章字符易被漏检、误检等问题,提出基于TextSnake的印章字符检测算法,记作TextSnake-CR算法。首先在特征融合中嵌入手工设计的感受野增强模块,使浅层特征拥有更大感受视野,从而有效地降低误检率;其次提出颜色特征提取模块,用于提取印章字符中颜色特征,增强模糊印章字符的检测精度;最后改进模型的分类损失函数,抑制背景噪声对模型的干扰,进一步提高模型检测性能。实验结果表明,TextSnake-CR在公开印章数据集与自制数据集上的F值分别达到90.71%和81.79%,与其他算法相比,有效地提高了印章字符检测准确率。Stamp characters are easy to be wrongly detected and missed due to fuzzy characters,variable aspect ratio,overlapping with background text and other factors.In order to solve these problems,a stamp character detection algorithm based on TextSnake is proposed,which is referred to as the TextSnake-CR algorithm.Firstly,a hand-designed receptive field enhancement module is embedded in the feature fusion network,so that the shallow features have a larger receptive field of view,and then effectively reduce the false detection rate.Secondly,the color feature extraction module is proposed to extract the color feature of the stamp character,and enhance the detection accuracy of the fuzzy stamp character.Finally,the classification loss function of the model was refined to suppress the interference of background noise on the model,and further improve the detection performance of the model.The experimental results show that the TextSnake-CR achieves the F score of 90.71%and 81.79%on the public and the homemade stamp datasets,respectively.Compared with other algorithms,the TextSnake-CR effectively improves the detection performance of stamp character.

关 键 词：印章字符检测 TextSnake 颜色特征提取 感受野增强 分类损失函数 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]