基于改进KNN-SVM的车辆图像光照检测模型

为了准确检测出车辆交通图像的光照类型，从而有针对性地矫正不同光照以减少其对车牌定位的影响，提出了一种基于改进K近邻和支持向量相融合（KNN-SVM）的车辆图像光照检测方法。首先融合了HSV空间亮度特征、灰度直方图特征和投影直方图特征作为车辆图像的光照特征，然后改进传统KNN-SVM中距离计算方法，定义为每类待检测样本到属于该类支持向量的距离，并在采集的全天候不同光照车辆图像上进行检测验证。实验表明，改进KNN-SVM将阈值获取时间提前，避免了传统KNN-SVM对超平面附近样本先SVM检测再KNN检测的重复检测，不仅降低了算法复杂度和运行时间，且检测准确率高于传统KNN-SVM和单独使用KNN或SVM时的值，最高达到了99.67%。

Vehicle image illumination detection model based on improved K Nearest Neighbor and Support Vector Machine

In order to accurately detect the light type of vehicle traffic image, so as to correct different lighting to reduce its impact on the license plate positioning, a vehicle image lighting detection method based on improved K Nearest Neighbor and Support Vector Machine（KNN-SVM） is proposed. Firstly, the HSV spatial brightness feature, the gray histogram feature and the projection histogram feature are fused as the light feature of the vehicle image, and then the distance calculation method is improved in traditional KNN-SVM, which is redefined as the distance between each class of samples to the class of support vectors, and the testing and verification is performed on the collection of all-weather different light vehicle image. Experiments show that, the improved KNN-SVM advances the time to threshold acquisition, avoids repeated detection of traditional KNN-SVM for SVM detection and KNN detection in the vicinity of the hyperplane, it not only reduces the algorithm complexity and running time, but the detection accuracy is higher than the traditional KNN-SVM and the value of KNN or SVM alone, which is up to 99.67%.