摘要
针对自然场景文本检测面临的文本形状差异大、场景复杂干扰多等诸多挑战,提出了一种基于像素聚合的自然场景文本检测模型。首先,设计了上采样和长短跳跃的嵌套巢式连接的特征融合模块,通过融合残差网络ResNet18提取的多尺度、多阶段的特征,增强网络特征提取的能力;其次,基于聚类的思想,引入像素聚合约束外围像素与文本中心区域的距离,实现复杂自然场景下的任意形状文本描述;最后,通过轻量级文本检测头实现像素级的字符分割,提高模型的效率。在ICDAR2015、CTW1500以及构建的工业字符数据集上对所提模型进行验证,结果表明该模型能胜任复杂自然环境下的文本检测任务,且在检测精度和检测效率上均优于现有先进文本检测器。
Aiming at numerous challenges faced in natural scene text detection,such as significant variations in text shapes and multiple interferences in complex scenes,a pixel-aggregation-based natural scene text detection model is proposed.Firstly,a feature fusion module with nested concatenation of upsampling and long-short skip connections is designed to enhance the network′s feature extraction capability by fusing multi-scale,multi-stage features extracted from the ResNet18 residual network.Secondly,inspired by clustering,a pixel aggregation constraint is introduced to minimize the distance between peripheral pixels and the text′s central area,enabling the description of text in arbitrary shapes within complex natural scenes.Finally,a lightweight text detection head is employed to achieve pixel-level character segmentation,thus improving the model′s efficiency.The proposed model is validated on the ICDAR2015,CTW1500,and a constructed industrial character dataset.Results demonstrate that the model is capable of handling text detection tasks in complex natural environments,and outperforms existing state-of-the-art text detectors in terms of detection accuracy and efficiency.
作者
张华东
钟羽中
涂海燕
佃松宜
ZHANG Huadong;ZHONG Yuzhong;TU Haiyan;DIAN Songyi(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)
出处
《组合机床与自动化加工技术》
北大核心
2024年第11期13-17,23,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家重点研发计划项目(2018YFB1307401)。
关键词
特征融合
像素聚合
文本检测
字符分割
feature fusion
pixel aggregation
text detection
character segmentation
