结合纹理分析和支撑矢量机的汽车牌照定位研究

License plate locating using support vector machines and texture analysis

为了提高汽车牌照定位的准确率,将拍摄到的汽车图像先作离散余弦变换(DCT),然后对频率系数量化,从量化值中提取图像的方向性,并将量化值的统计特性和图像的方向性结合起来构造频域特征矢量;在空域中提取图像的角二阶矩、对比度、相关性和熵4个特征量,构造空域特征矢量.结合支撑矢量机,在粗分类时找到牌照候选区域,在细分类时从牌照候选区分离出真实的牌照.分类过程综合考虑分类速度、分类准确率和感兴趣的区域,可采取灵活的分类方式.实验表明,该方法对于汽车牌照定位具有较好的效果,可操作性较强.

To improve the accuracy of plate locating, the primary car image was transformed using discrete cosine transform （DCT）, and qualification values of DCT coefficients were obtained. From these values, the statistical and directional features of image were picked up to construct a vector in frequency domain. Four features of spatial domain-angle second moment, contrast, relativity and entropy were used to constitute another vector. With these vectors, license plate candidate area was found by coarse SVM classification and the license plate was located by fine SVM classification. During classification, problems of speed, precision, area of interest and a flexible step were taken into consideration. The experiment results showed that the way suggested in this paper is good for license plate locating and very operable.