自然场景图像中维吾尔文检测算法  

Uyghur alphabets detection algorithm in natural scene images

作　　者：王德青 吾守尔·斯拉木[1,2] 许苗苗 WANG De-qing;Wushouer·Silamu;XU Miao-miao(College of Information Science and Engineering,Urumqi 830046,China;Xinjiang Key Laboratory of Multilingual Information Technology,Urumqi 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046 [2]新疆大学新疆多语种信息技术重点实验室,新疆乌鲁木齐830046

出　　处：《东北师大学报（自然科学版）》2022年第3期76-82,共7页Journal of Northeast Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61433012);国家重点研发计划项目(2017YFC0820702-3)。

摘　　要：针对自然场景图像中维吾尔文检测难度较大、字体粘连、形状较小等问题,提出了改进的DBNet算法.该网络结构主要由拆分注意力网络ResNeSt50和特征金字塔网络FPN融合进行特征提取,然后对提取的特征进行预测,之后经过DB操作获得包含丰富语义信息和位置信息的维吾尔文字检测结果.经过在数据集的实验测试与对比,所提出的算法考虑了语义信息和位置信息对检测结果的影响,得到的准确率达到了0.767 2,F值达到了0.67,有效地提高了检测的准确率.An improved DBNet algorithm is proposed to solve the problems of difficult detection, font conglutination and small shape.The structure of DBNet is mainly extracted by the fusion of divided attention network ResNeSt50 and feature pyramid network FPN,then the extracted features are predicted, and the Uyghur text detection results containing rich semantic information and location information are obtained by DB operation.After testing and comparing the data set in this paper, the algorithm considers the influence of semantic information and position information on the detection result, the accuracy rate is 0.767 2,the F value is 0.67,and the detection accuracy is improved effectively.

关 键 词：拆分注意力 维吾尔文检测 特征金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]