注意力机制与自适应尺度融合的场景文本检测  

Scene text detection based on attention mechanism and adaptive scale fusion

作　　者：梁雨欣 毕晓君 LIANG Yuxin;BI Xiaojun(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China;School of Information Engineering,Minzu University of China,Beijing 100081,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]中央民族大学信息工程学院,北京100081

出　　处：《应用科技》2023年第1期45-50,111,共7页Applied Science and Technology

摘　　要：在场景文本检测任务中,由于图像背景复杂、文本实例尺度不一等问题,导致现有模型的文本检测精度不高。为此,本文设计了一种基于注意力机制与自适应尺度融合的场景文本检测模型。首先,通过引入高效通道注意力机制,提高了特征提取网络的表征能力,降低了文字的漏检率和误报率;其次,通过设计自适应尺度融合模块,动态融合不同尺度特征,增强了模型对不同尺度文本实例的检测和定位能力。实验结果表明,本文提出的模型在Total-Text和MSRATD500共2个数据集上的F综合指标分别达到85.1%和84.1%,在同类型算法中处于领先水平。At present,due to complex image background and different scales of text instances in scene text detection tasks,the text detection accuracy of existing models is not high.Aiming at this problem,we design a scene text detection model based on attention mechanism and adaptive scale fusion.The introduction of high-efficiency channel attention mechanism improves the representation ability of feature extraction network and reduces the false positive rate and false negative rate of texts.At the same time,an adaptive scale fusion module is designed to dynamically fuse the features of different scales,which enhances the detection and localization ability of the model for text instances of different scales.The experimental results show that the comprehensive index of F in Total-Text and MSRA-TD500 datasets is up to 85.1%and 84.1%,respectively,which proves its leading level in the same type of algorithms.

关 键 词：场景文本检测 任意形状文本 注意力机制 自适应尺度融合 概率图 向心偏移图 深度学习 自然场景 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]