Research And Implementation Of Vehicle Feature Recognition System Based On Deep Learning

With the continuous improvement of urbanizational level in our country, more and more vehicles are serving our daily life. On the other hand, It can also result in some problems, such as the low efficiency of artificial management, how to process traffic violations quickly and accurately, and so on. The essence of the above problems is how to identify the vehicle feature, and then combine with the Internet technology to enhance the managemental efficiency, and ultimately achieve the intelligent management of vehicles.Vehicle feature recognition is a research focus in the field of target detection and recognition in recent years, especially the vehicle license plate recognition. But the product in the market is not universal, especially under the condition of uneven illumination, license plate tilt and so on. Therefore, It has great application value to do further research on the vehicle feature recognition system.Based on the above background, this thesis will do deeply further research on license plate location, tilt correction, character segmentation and character recognition,color of vehicle and other modules, in feature recognition system of vehicle, combine with the latest computer vision technology, and then put forward our own methods and improvement, carry out experimental comparative analysis, and ultimately achieve good results.The main contexts of this thesis are as follows:1, vehicle license plate location algorithm: First, the collected vehicle image will be gray-processed, calculated horizontal gradient. Then according to the theory that horizontal gradient information of the license plate area is rich in the image, we get rough candidate regions for precise positioning. The candidate regions which do not conform to the inherent geometric features of the license plate are excluded. Finally, an SVM classifier based on HOG features are trained and used for screening, to get the final candidate regions of license plate.2, tilt correction algorithm: This paper proposed an algorithm which based on row(column) projection difference statistics to solve the problem of horizontal and vertical tilted plate, and accurately detect the tilt angle, finally achieve a good correction effect.3, character segmentation algorithm: Based on the traditional vertical projection algorithm and combined with the distribution characteristics of license plate character, We propose the special six-character license plate segmentation algorithm, finally segment the Chinese character in the license plate.4, character and color of vehicle recognition: We introduce the remarkable convolutional neural network in classification problems in deep learning. In view of the Chinese characters, letters and numbers, color of car,We design different network structures to avoid the artificial feature extraction process in the traditional method and improve the robustness of recognition.All algorithms in this thesis are developed using C++ language based on OpenCV1.0, with the development tool of Microsoft Visual Studio 2010.