Abstract:

Character recognition accuracy is directly affected by character segmentation. However, in the process of pressed characters in low quality images with uneven brightness, due to the uneven distribution of image brightness and the low contrast between the target characters and the background region, the accurate segmentation of pressed characters is difficult. The traditional character segmentation methods based on the binary images are difficult to determine the optimal image-binary threshold when dealing with the characters in the above situation. Therefore, an iterative closed-loop feedback segmentation method based on the evaluation function of segmentation effect is proposed. By using the evaluation function to evaluate the segmentation effect, it can judge whether to repartition or not. In addition, to improve the efficiency and accuracy of segmentation, weighted smoothing filter is used to remove the distorted waveforms in the grayscale waveforms. The generalized learning vector quantization algorithm is used to determine the optimal filter weight factor. Then, the waveform variation trend is analyzed to determine the character segmentation position. Experimental results show that the segmentation accuracy of the algorithm can reach 98.5%.

Key words: characters segmentation, uneven brightness images, segmentation effect evaluation function, iterative closed-loop feedback, generalized learning vector quantization algorithm

摘要：

字符分割效果，直接影响识别精度。在处理亮度不均匀低质量图像中的压印字符时，由于亮度分布不均匀、目标字符与背景区域对比度较低，传统基于二值化图像的字符分割方法在处理上述情况下的压印字符时，难以确定最优二值化阈值，分割准确率较低。提出基于分割效果评价函数的迭代闭环反馈分割方法，通过建立评价函数对分割效果进行评估，以提高分割效率及准确率。借助加权平滑滤波，去除灰度波形图中的畸变波形；并利用广义学习矢量量化算法，确定最优滤波权重因子；通过分析波形变化趋势，确定字符分割位置。实验结果表明，该算法在批量处理亮度不均匀低质量图像中的压印字符时，分割准确率可达98.5%。

关键词: 字符分割, 亮度不均匀图像, 分割效果评价函数, 迭代闭环反馈, 广义学习矢量量化算法