多层次MSER自然场景文本检测

Natural scene text detection based on multi-level MSER

提出一种新的基于多层次最大稳定极值区域(MSER)的自然场景文本检测方法,其由候选区域的提取和文本检测组成.在候选区域提取过程中,采用多层次MSER区域提取方法:通过对原始图像进行多个颜色空间变换和多尺度放缩得到多个变换后的图像,采用多个阈值对其进行MSER区域检测,并将检测到的区域作为候选区域用于文本检测.检测过程中,对候选区域提取手工设计的底层特征和基于卷积神经网络(CNN)的深层特征,训练一个随机森林回归器对特征进行分类得到字符区域,再将其合并成单词区域,并进行相似的特征提取和分类,从而得到最终的文本检测结果.使用2个标准的数据库(ICDAR2011和ICDAR2013)对提出的方法进行性能评价,F指标在ICDAR2011和ICDAR2013上均为0.79,表明了所提出的自然场景文本检测方法的有效性.

A novel scene text detection method based on multi-level maximally stable extremal regions（MSER）was proposed,which consisted of two main stages,including candidate regions extraction and text regions detection.In the stage of candidate regions extraction,a multi-level MSER region extraction technique was developed by considering multiple color spaces,multiple scale transformations of original image and multiple thresholds of MSER detection.All extracted regions from the input image were used as candidate character regions for text region detection.In the stage of text detection,the hand-designed bottom features and CNN based features were extracted for each candidate character region as first,then a random forest regressor trained from training datasets was used to get the character regions.After that,the character regions were merged to form candidate word regions,from which the features were extracted and classified to get the final text detection results by using the similar process of candidate character region classification.The proposed method was evaluated on two standard benchmark datasets,including ICDAR2011 and ICDAR2013,and both got the F-measure performance of 0.79,respectively,Which demonstrates the effectiveness of the proposed natural scene text detection method.