基于多特征检测与支持向量回归的图像文本提取算法

The text extraction algorithm based on multi-feature detection and support vector regression

为解决复杂背景中难以有效提取场景文本的问题,提出了一种基于多特征检测与支持向量回归的图像文本提取方案。为有效区分文本与非文本边缘,基于图像边缘,提取场景中三个文本特征。将得到的三个文本特征进行多尺度融合,利用文本融合特征检测候选文本边界,有助于检测不同大小的文本,提高对不同类型的图像退化的鲁棒性。对于每个检测到的候选文本边界,根据邻域窗口中的像素来估计每个像素的局部阈值,利用局部阈值自适应分割提取候选字符。引入支持向量回归模型对文本像素与图像背景精确分离,消除非文本边界,提取真实字符和单词。实验表明：与当前文章提取技术相比,所提方法具有更好的鲁棒性,能适用各种变化的复杂场景文本提取,具有更优的PrecisionRecall曲线与F测量值。

It is difficult to extract scene text from complex background,and a text automatic detection and extraction scheme based on multi-feature detection and support vector regression was proposed.In order to distinguish the text from the non text edges effectively,three text features were extracted based on the image edge.And the three text features by multi-scale fusion,feature fusion using text detection candidate text boundaries helps detect text with different sizes of different types of degraded image to improve the robustness.For each candidate text boundary detected,the local threshold of each pixel is estimated according to the pixels in the neighborhood window,and the candidate characters are adaptively segmented by local threshold.The support vector regression（SVR）model is introduced to separate the text pixels from the image background,eliminate the non text boundaries and extract the real characters and words.Experiments show that the proposed method has better robustness and can be applied to text extraction in various complex scenes.Compared with other algorithms,this algorithm has better Precision-Recall curve and F measurement value.