| 引用本文: | 杨帅磊,李岐龙,陈杰,凡高娟,张重生.基于特征过滤和自适应融合机制的不规则场景文本检测算法[J].哈尔滨工业大学学报,2025,57(10):165.DOI:10.11918/202212061 |
| YANG Shuailei,LI Qilong,CHEN Jie,FAN Gaojuan,ZHANG Chongsheng.Irregular scene text detection based on feature filtering and adaptive fusion mechanisms[J].Journal of Harbin Institute of Technology,2025,57(10):165.DOI:10.11918/202212061 |
|
| 摘要: |
| 自然场景图像中的文本存在背景复杂、形状各异、方向多样、光线多变等特点,为提升场景文本尤其是不规则场景文本的检测性能,提出一种基于特征过滤和自适应特征融合机制的不规则场景文本检测网络FGANet（feature guided adaptive network）,其特征聚合引导模块通过空洞卷积扩大感受野、提升网络的特征表达能力,其自适应特征融合模块能够将深层的语义信息与浅层的细节信息进行综合利用,使算法获得更强的文本感知能力。结果表明,在场景文本检测方面,FGANet相较于对比算法,在ICDAR2015、CTW1500、MSRA-TD500和TotalText 4个基准数据集上的F值分别提升了2.4%、1.3%、1.8%和1.4%,性能提升较为显著。 |
| 关键词: 自然场景图像 场景文本检测 特征过滤 自适应特征融合 文本感知能力 |
| DOI：10.11918/202212061 |
| 分类号:TP183 |
| 文献标识码:A |
| 基金项目:教育部人文社会科学研究一般项目(23YJAZH210)；河南省科技攻关项目(232102211021) |
|
| Irregular scene text detection based on feature filtering and adaptive fusion mechanisms |
|
YANG Shuailei1,LI Qilong1,CHEN Jie1,FAN Gaojuan1,ZHANG Chongsheng1,2
|
|
（1.Henan Key Lab of Big Data Analysis and Processing(Henan University), Kaifeng 475001, Henan, China;2.Lab of Yellow River Heritages(Henan University), Kaifeng 475001, Henan, China）
|
| Abstract: |
| Text in natural-scene images often present characteristics of complex backgrounds, varied shapes, multiple orientations and changing illumination. In order to improve detection performance for scene text, particularly irregular text, we propose the feature guided adaptive network (FGANet), an irregular-scene text detection network based on feature filtering and adaptive fusion mechanisms. In specific, FGANet designs a module that utilizes dilated convolution to enlarge the receptive and enhance the network′s feature representation capability. Its adaptive feature fusion module integrates deep semantic information with shallow detailed information, enabling stronger text-awareness. Experiments results show that for scene text detection, FGANet achieves notable improvements in F-score over comparative methods on four benchmark datasets: ICDAR2015, CTW1500, MSRA-TD500, and TotalText, with gains of 2.4%, 1.3%, 1.8%, and 1.4%, respectively. |
| Key words: scene text images scene text detection feature filtering mechanism adaptive feature fusion mechanism text perception ability |
