具有自校正与注意力机制相结合的场景文本检测

Scene text detection with self-calibration and attention mechanism

在日常生活中,存在着丰富的文本信息,对这些信息的提取,能够极大地提高人们的生活品质。但自然场景中文本信息表达形式丰富多样,文本形状各异,在检测过程中存在误检、文本区域定位不准问题。针对以上不足,本文提出了一种具有自校正与注意力机制相结合的文本检测方法。首先,在ResNet50骨干网络中嵌入自校正卷积(self-calibrated convolution, SConv)及高效通道注意力(efficient channel attention, ECA),使网络能够校正全局无关信息的干扰,并集中关注于文本区域,提取更加丰富的语义信息;其次,在特征融合后加入协调注意力(coordinate attention, CA),纠正不同尺度的特征图在融合过程中产生的位置偏差。最后,通过修正后的特征图预测得到多个不同尺度的文本实例,采用渐进尺度扩展算法,求出最终检测到的文本实例。实验结果表明,在任意方向数据集ICDAR2015以及弯曲文本数据集Total-Text、SCUT-CTW1500上,相比于改进前的ResNet50综合指标F值分别提升了1.0%、5.2%、5.4%,证明了本方法具有良好的检测能力。

In daily life, there are rich text information, the extraction of such information can greatly improve people′s quality of life.However, there are various forms of text information expression and different text shapes in natural scenes, which result in false detection and inaccurate location of text regions.In order to solve these problems, this paper proposes a text detection method with self-calibration and attention mechanism.Firstly, the self-calibrated convolution(SConv) and efficient channel attention(ECA) are embedded in the backbone of ResNet50 to correct the interference of irrelevant global information and concentrate on the text area to extract more abundant semantic information Secondly, coordinated attention(CA) is added after feature fusion to correct the position deviation of feature map in different scale.Finally, several text instances of different scales are predicted by the modified feature map, and the final detected text instances are obtained by using the progressive scale expansion algorithm.The experimental results show that the comprehensive index F-measure is increased by 1.0%,5.2% and 5.4% respectively compared with the unmodified ResNet50 on the arbitrary direction data set ICDAR2015 and the curved text data set Total-Text and SCUT-CTW1500.It is proved that this method has good detection ability.