基于机器学习的多车牌识别算法应用研究

Research on Application of Multiple License Plate Recognition Algorithm Based on Machine Learning

为满足车牌识别系统中对多车牌识别的高准确率的应用需求,提出一种基于优化参数的SVM(支持向量机)进行车牌定位。考虑到雾霾天气下拍摄的图片对车牌识别的影响,先进行去雾处理,然后使用Sobel算子对图片进行垂直边缘检测,结合形态学处理确定候选车牌轮廓,对候选车牌轮廓通过外接矩阵的长宽比初步判断符合车牌的区域。加上国内车牌的颜色单一,结合HSV颜色模型定位,很大程度上提高了车牌的定位率。对定位出的候选车牌区域进行训练和SVM模型判断,确定出符合车牌的区域。同时对车牌进行字符分割后,使用单独为车牌汉字训练的ANN模型进行字符识别。此外针对不同的场景提供训练模式,系统可以训练特定场景下的SVM模型。经验证,该系统能够满足多车牌识别的实际应用,鲁棒性和准确率相比通用模型提高近20%。

In order to meet the requirements about high accuracy of multi-license plate recognition for license plate recognition system,we propose a SVM（ support vector machine） based on optimization parameters to locate the license plate.Taking into account the impact of the images taken in hazy weather on the license plate recognition,the first thing is to remove the image haze.Then,the Sobel operator is used for vertical edge detection.According to the morphological processing,the candidate license plate contour is determined and the area of license plate is judged initially by the length and width of the external matrix to the candidate license plate contour.In addition,the color of the domestic license plate is single,so the localization rate of the license plate is improved to a great extent by combining the localization of HSV color model.The SVMmodel is trained to judge the candidate license plate and determine the area which conforms to the license plate.At the same time,after the character segmentation of the license plate,the character recognition is carried out by using the ANN model which is trained by the Chinese characters for the license plate.In addition,the system can train the SVMmodel in a specific scenario in order to provide a training pattern for different scenarios. It is verified that this system can meet the practical application of multi-license plate recognition and its robustness and accuracy increases nearly 20 percent compared to the general model.