基于可变形部件模型的车牌检测

Plate Detection Based on Deformable Part Model

针对传统的利用车牌先验信息后检测的方法受背景、分辨率、光照和车辆姿态影响较大的问题,提出了一种基于可变形部件模型的车牌检测方法。该方法首先提取车牌样本的HOG特征,分别获取根滤波器模型和部件滤波器模型,融合后得到最终的检测模型;然后计算待检测图像的HOG特征,用检测模型分别计算根滤波器的得分和部件部件滤波器的得分;最后将根滤波器的得分与进行距离变换之后的部件滤波器的得分相加,根据相加后的结果确定车牌的真实位置。实验结果表明,该方法能够准确检测出车牌位置,能应用于各种复杂背景、低分辨率、恶劣光照和任意车辆姿态的环境中,具有一定的实用价值。

While the traditional plate detection methods based on the license plates＇ prior information are greatly influenced bythe background, resolution, light and vehicle attitudes, a license plate detection method based on deformable part model is presented.First, the method extracted the license plates＇ HOG features, and used the features to train a root filter and several part filters, and thengot a detection model. Second, the method calculated the HOG features of images to be detected, and used the detection model to calculatethe root filter＇s score and the part filters＇ score separately; Finally, the method processed the part filters＇ score based on distancetransform, and then added to the root filter＇s score, and the real plate position was gained based on the final score. The experimental resultsshow that the method can accurately gain the license plate＇s location, and can be applied to various scenarios such as complexbackground, low resolution, bad light and any vehicle attitudes, so the method has a certain practical value.