一种基于字符组合的复杂环境车牌检测方法

A License Plate Detection Based on Character Combination in Complex Scenes

多变的自然环境如光照变化、天气变化等,以及复杂的监控场景如拍摄背景、拍摄距离、拍摄角度、设备像素等因素,严重影响了车牌检测算法的准确性与可靠性。环境的复杂性使得车牌定位变得更加困难,传统的基于车牌定位的检测与识别方法识别率不高。新型的基于字符组合的车牌检测与识别方法,跳过车牌定位过程转为直接定位字符,对复杂背景以及光照强度变化等具有较好的鲁棒性。但是基于字符组合的车牌检测与识别方法也有其不足,识别准确率对字符分割具有更强的依赖性,且对于字符形变的鲁棒性低。对此,文中分别提出了一种基于多粒度区域生长的字符分割算法和一种基于多特征聚类的字符组合检测算法。实验结果表明,多粒度区域生长分割得到的目标字符区域及其稳定粒度区间,很好地保证了分割的可靠性,通过多粒度决策融合提高了字符识别率;利用特征子集进行字符组合聚类分析,对字符形变具有良好的鲁棒性。

Changeable natural environment like illumination intensity and weather changes,and complex surveillance scenes like background,distance,shooting angle and device pixel,all these factors seriously affect the accuracy and reliability of the license plate detectionalgorithm. The complexity of the scenes makes it more difficult to locate the license plates,which causes the low recognition rate of traditional license plate detection and recognition method based on locating plate. A new license plate detection and recognition method basedon character combination locate the plate characters directly instead of license plate itself,which has a better robustness to complex background and illumination intensity change. While the new method also has shortcomings due to its stronger dependence on character segmentation and low robustness to character deformation. According to the shortcomings,we propose a character segmentation algorithmbased on multi-grain region growing and a character combination detection algorithm based on multi-feature clustering. Experimentshows that the target character regions and their stable granularity interval ensure the reliability of segmentation and the accuracy of identification via multi-grain decision fusion. Also,using feature subsets to do character combination clustering has a good robustness to character deformation.