基于深度学习的倾斜车牌矫正识别

Identification of Tilting License Plate Correction Based on Deep Learning

车辆号牌矫正识别技术是交通系智能化体系中一项非常重要的基础功能。拍摄的车牌图像受到各种外部环境的影响而导致图像质量受损严重,文中通过运用深度学习的霍夫变换改进算法,针对大角度倾斜的车辆号牌研究其矫正识别技术。首先,利用基于深度学习的特征提取和识别方法,解决车辆号牌中相似字符的分类问题。建立五层深度网络,通过由底层向高层的顺序,获取归一化处理后的图像中相近字符特征,然后通过卷积函数提取出字符边缘特征并完成差异性比较;其次,利用集成深度学习的不同层次的识别方法有效解决车牌图片字符的识别效率问题。利用支持向量机完成对字符的统计和分类工作,通过卷积神经网络对其进行识别优化,分析模糊相近字符并最终获取车辆号牌识别图像。最后,在充分利用区域范围分划和霍夫变换的车牌倾斜校正方法后完成大角度车牌的矫正和识别。实验证明,采用文中方法能够对百幅以上大角度倾斜车牌图像实现有效矫正与识别。

Vehicle plate correction and recognition technology is a very important basic function in intelligent system of traffic department. The image quality of the license plate image is seriously damaged due to the influence of various external environments. In this paper, we use the hough transform to improve the algorithm, and study the correction and recognition technology of the vehicle plate with large angle. Firstly, the feature extraction and recog-nition method based on deep learning is used to solve the classification problem of similar characters in vehicle number plates. A five-layer depth network is established to obtain the similar character features in the image after normalization processing through the order from the bottom to the top, and then extract the character edge features through convolution function and complete the difference comparison. Secondly, the recognition efficiency of li-cense plate image characters is effectively solved by using different recognition methods of integrated depth learn-ing. Support vector machine is used to complete the statistics and classification of the characters, which is identified and optimized through convolutional neural network, and the fuzzy close characters are analyzed and the vehicle li-cense plate recognition image is finally obtained. Finally, the correction and recognition of large angle license plates are accomplished by using the method of region division and hough transform. Experimental results show that this method can effectively correct and recognize over 100 large Angle tilting license plate images.