基于笔画宽度特征和剪枝算法的自然场景标签检测

针对自然场景标签图像背景复杂、目标区域形式多样、色调单一、空间集中等问题,提出了一种基于笔画宽度特征和剪枝算法的自然场景标签检测方法。根据图像的特点,首先利用阈值法和启发式规则,初步筛选出字符候选区;然后通过设计的笔画宽度特征提取算法,获得每一个候选区的文本相似度,融合基于惩罚函数的剪枝算法实现背景区域的滤除,得到进一步分割后的标签检测区域;使用形态学处理和区域面积的细分割后,最终生成目标检测图像。多组对比实验检测结果表明,本文算法具有良好的目标检测效果和优异的普适性。

Label detection based on stroke width feature and pruning algorithm in natural scene images

In this paper,a label detection method in natural scene images based on stroke width feature and pruning algorithm is proposed to solve the problems of complex background,diverse target forms,single tone and spatial concentration. Firstly,it uses threshold method and heuristic rules to select candidate regions preliminarily according to the characteristics of processed image. Text similarity of each candidate region is obtained by stroke width feature extraction algorithm designed in this paper. The further label detection regions are segmented by filtering of background regions realized by pruning algorithm based on penalty function. Using morphological processing and fine division of region area,the target detection image is generated eventually. Multiple groups of comparative experiments show that this method has good effect on object detection and excellent universality.