| The License Plate Recognition (LPR) system is a significant sector of the modern intelligent traffic management, which has a far-reaching future for the application of traffic management and police security, and attracted by all over the world at the present.Five key parts of license plate recognition system have been researched, which are image preprocessing, license plate location, image binarization, character segmentation and character recognition.Ⅰ.Image preprocessingTraditional methods of image enhancement like gray stretch and histogram equalization have been analyzed. Considering the deficiencies of these global enhancement algorithms, such as noise over-enhancement, a method of local enhancement algorithm based on section thinking has been put forward. It enhances the regions where are similar to the license plate's, and it can be not only has a significant enhancement to the region of license plate but also with fewer background noise that it was introduced. Moreover, it will be especially effective for the low contrast vehicle images which need local enhancement.Ⅱ.License plate locationAiming at low precision of license plate location in the color vehicle images with complex background, a hybrid license plate location method based on edge detection and color features has been proposed.License plate region contains rich edge information, so firstly, the vertical edges are extracted out, and then most of the background and noise edges are removed, after then two candidate regions are ascertained by vertical edges statistical analysis, and finally the region of real plate is filtrated by color features of plate .The method is accurate and robust.ⅢImage binarizationAiming at the characteristic of license plate image with complicated-background,a binarization algorithm of license plate image based on improved compensation-based binzrization has been proposed.Firstly,the quality of plate image is enhanced by homomorphic filtering for eliminating the negative of the nonuniform illumination, and then the plate image is binaried by an improved compensation-based algorithm. Some influences which come from the nonuniform illumination can be surmount robustly, and the effect of the binarization is well.Ⅳ.Character segmentationConsidering the influence of noise and some unavoidable slants in the area of located license plates, an approach of rectifying distorted license plate image based on character feature has been presented to rectify the binary plate images directly. And then a method of character segmentation based on vertical projection and prior knowledge of license plate has been used to segment single character from rectified plate image. The problems of rupture, adhesion and geometric distortion existed in the process of character segmentation can be solved effectively, and the method has high segmentation accuracy.Ⅴ.Character recognitionAccording to the particularity of domestic license plate character, a classification technology by combination of feature extraction and BP neural network algorithm has been adopted. The feature of rough grid has been chosen as the feature of character recognition, and directly input the improved feature of rough grid to BP neural network classifier to recognize the license plate character. Aimed at the confusable characters, a careful neural network classifier has been designed for distilling the detail features of characters. Considering the phenomena of stroke conglutination found in Chinese characters and because of some province characters which are complex in structure and compact in strokes, some representative stroke conglutination stylebooks are suitably added to the network training stylebooks of Chinese character based on standard stylebooks. The ratio of character recognition is improved efficiently by the method.The algorithm above-mentioned has been actualized by experiments. The experimental results show that the algorithm has some advantage such as accurate location, high recognition rate, and high utility value. |
PDF Full Text Request