多尺度积角点检测和视觉颜色特征的鲁棒车牌定位算法

车牌图像包含的尺度、仿射变化及其复杂的背景是影响车牌定位准确度的重要因素。在高斯差（DOG）尺度空间框架下，笔者提出了一种基于多尺度乘积的角点特征和视觉颜色特征提取及其相融合的车牌定位算法。基于高斯差尺度空间的图像边缘信息，应用多尺度乘积分别提取具有尺度和仿射不变特性的角点和颜色特征，并在两特征融合结果基础上确定车牌位置候选区域；最后通过车牌区域特征点之间的距离及密集关系实现车牌的准确定位。对大量实拍的复杂环境下的车辆图像进行测试表明，该算法对车牌定位具有快速、高效的定位效果，且在噪声、仿射变换等方面的鲁棒性表现较好。

A robust license plate location algorithm based on multi-scale feature fusion corner detection and visual color features

Scale, affine variance and complex background are important factors affecting the accuracy of license plate location. We presented a method of license plate location based on multi-scale product of corner detection and visual color features in difference of Gaussian (DOG) scale space. Based on the image edge information in DOG scale space, we first extracted scale-and-affine-invariant corner and color features through multi-scale multiplication, and then obtained the candidate license plate location by fusing the corner and the color features. Finally, we accurately located the license plate by using the distance between the feature points in the plate region and the intensive relationship of the points. Experiments on several real-world vehicle image data sets under complex conditions have verified the proposed method has high effectiveness and efficiency in locating license plates, and greater performance in robustness of noise and affine variance than other state-of-art license plate localization methods.