基于MSER的自适应学习自然场景文本检测

Adaptive Learning Text Detection in the Natural Scene Based on MSER

自然场景下的文本检测分为提取候选区域和滤除非文本两个阶段.在候选区域提取阶段,针对最大稳定极值区域(Maximally Stable Extremal Regions,MSER)算法对噪声、模糊敏感,检测性能不高的问题,提出改进的MSER算法,首先通过梯度图增强图像字符边缘,然后采用MSER算法提取文本区域,最后利用多机制抑制策略进行粗过滤.在非文本滤除阶段,针对候选域中非文本区域过滤不彻底的问题,提出基于SVM的多特征自适应权值融合的非文本滤除算法,首先对标识样本库提取HOG、统一化LBP、颜色感知差异(Color Perception Difference,CPD)特征,使用提出的权值计算公式自适应分配权重融合三种特征,然后采用粒子群算法寻找SVM最优参数训练分类器,最后将候选区域送入训练好的分类器滤除非文本.实验结果表明,改进的文本检测算法能够达到理想的检测效果.

Text detection in natural scenes is divided into two stages:extracting candidate regions and filtering out text.In the candidate region extraction stage,for the problem that the Maximally Stable Extremal Regions(MSER)algorithm is sensitive to noise and blur,and the detection performance is not high,an improved MSER algorithm is proposed.Firstly,the edge of the image character is enhanced by the gradient map,and then The MSER algorithm is used to extract the text region.Finally,multi-mechanism suppression strategy is used for coarse filtering.In the non-text filtering stage,the non-text filtering algorithm based on SVM multi-feature adaptive weight fusion is proposed for the problem of incomplete filtering of non-text regions in the candidate domain.Firstly,the HOG is extracted from the identification sample library,and the LBP is unified.Color Perception Difference(CPD)feature,using the proposed weight calculation formula to adaptively assign weights to fuse three features,then use particle swarm optimization to find the SVM optimal parameter training classifier,and finally sendthe candidate region into training.The classifier filters out the text.Experimental results show that the improved text detection algorithm can achieve the desired detection effect.