基于DPM的自然场景下汉字识别方法

自然场景下, 汉字背景复杂且形态各异, 导致传统识别方法中的文本定位与文本矫正过程难以进行。为了避免这些问题, 采用物体识别方法中的可变部件模型（DPM）进行识别。该方法将汉字视为物体类, 训练其对应的参数模板, 然后采用滑动窗口的方法遍历待检测图片, 以判断图片中是否存在目标汉字。实验表明, 该方法对简单独体汉字有较好的检测效果, 但对于多笔画复杂汉字, 由于模型自身结构特点, 效果并不明显。

Chinese characters recognition in natural scenes based on DPM

Chinese characters take on variable appearances in natural scenes, that makes some inevitable procedures in traditional detecting methods hard to carry on, such as the text location and text modification. In order to avoid such difficulties, this paper used DPM to detect Chinese characters in natural scenes. The method treated each character as an object class and trained the corresponding template. By using sliding windows, it traversed the detecting image to judge whether the image contained the target character. Experimental results show that the detecting effects of sole characters are satisfying and robust. As for the multi-stroke-characters, the effects aren't so obvious owing to the model's structural features.