面向复杂场景图像的文本定位新方法

A Novel Text Location Method for Complex Scene Pictures

针对复杂场景文本,提出了通过投影产生候选文本块的新算法和针对该算法的候选文本块分析方法。首先根据MLP网络的输出确定图像每个像素点是文本像素点还是非文本像素点,得到候选二值图像。然后根据候选二值图像使用投影法生成候选文本块,针对该投影法,本文提出了频率分析法剔除非文本块,有效的提高了定位准确率。实验表明,本文的方法实现简单,而且可以得到较好的文本定位效果。

This paper proposes a novel candidate text-block generating algorithm using projection for complex scene text and a novel method for that algorithm. First, text-pixels and nontext-pixels are discriminated based on the output of MLPs, and a candidate binary image is got. Then, based on candidate binary image, candidate text-blocks are generated using projection algorithm. We propose the frequency analysis method for projection algorithm to eliminate non-text blocks, that method increases location precision effectively. Experimental results show that, our approach is simple to implement and can get good text location result.