Abstract: To extract the characters in vehicle plate, which is a special scene image, a compressed-domain text extraction algorithm is proposed. It integrates text region detection, location, tracking and text recognition in a united framework. Since DCT coefficients can be directly utilized to represent the texture feature, a decision rule for adaptive threshold is defined based on the sum of weighted frequency components to detect the plate region. The text region is further located by the projection profile of DC+2AC texture intensity. The inter-frame text region tracking is completed by using the average motion vector of some macro-blocks, whose intersected area with located vehicle plate is larger than predefined value. The OTSU algorithm is adopted to convert the decompressed vehicle plate into binary image for OCR processing. Experiment demonstrates the effectiveness of the proposed approach. The recall and precision ratios can reach up to 95% and 96.2%, respectively. It is suitable for the text extraction of other type’s scene as well.

Key words: vehicle plate character extraction, compressed domain, Moving Picture Experts Group（MPEG）-2

摘要： 针对车辆牌照字符这一特殊场景文本的提取，提出一种压缩域文本提取算法，它集文本区域检测、定位、跟踪和字符识别等环节于统一的框架。直接利用DCT系数可以表征图像的纹理特征这一特性，建立加权频率分量和的自适应阈值判断规则初步检测车辆牌照区域，利用分块的DC+2AC纹理值投影特征进行文本定位；利用与定位得到的文本框相交面积高于预定阈值的宏块运动矢量均值实现跟踪；在OCR识别前采用OTSU算法把车辆牌照区域图像转换为黑白二值图像。实验表明了算法的有效性，召回率与准确率分别可达95%与96.2%，该算法也适用于其他类型文本的提取。

关键词: 车牌字符提取, 压缩域, 运动图像专家组标准（之一）（MPEG-2）