Japanese Vehicle License Plate Location, Segementation And Recognition

With the rising of people'living standard, automobile becomes household in millions families. To solve the increasingly serious problem of urban traffic congestion, intelligent transportation system has been proposed. Although it already has a lot of achievements and many products available, because of the variable recognition environment, there are still many scholars for further study.In this paper, I conduct deep study on the status of the license plate recognition system and existing technology and identify the research emphasis and difficulty of Japanese license plate recognition. The major research tasks including three parts: license plate location, vehicle license plate character segmentation and character recognition algorithm. In the license plate locating module, I propose a novel approach which employing HOG (gradient direction histogram) feature combining AdBoost classifier to address the lack of sunlight, excessive exposure, shadows, ghosting problems; then use connective component method as verification algorithm of the candidate areas to further improve location accuracy of system, because there are a few similar number connective components which have close aspect ratio and number of sum pixels. This approach has a good adaptability for the license plates have various backgrounds; binarization algorithm of license plate is taking Otus approach, for the condition of license plates with shadows, using an improved Bernsen binarization algorithm to solve; for the problem of image of Japanese vehicle license plate areas are blur, propose an approach which can enhance image of license plate to easily extract characters.In character segmentation section, for some low-resolution sample images(4mm/pixel) which captured by analog camera and the uneven illumination, this paper presents an approach which combine the HOG features and AdaBoost, this algorithm first locate figures in the license plate, then segment other characters according to the proportion of characters layout of normalized license plates.In character recognition part, I put forward three new character feature extraction methods: the diagonal density of strokes in the directions (DDSD), the diagonal density of strokes in the directions based on grid (DDSDG) and periphery feature of character based on grid (PCG). These feature extraction methods deeply explore the shape information of both inside and outside of the character and have a good effect in identifying address names and Hiragana; at the same time, I design a parallel multi-SVM recognition system which fusion many kinds features.?By 8283 Japanese automobile images experimental tests, the system overall recognition rate is 97.8%, while the correct location rate is 99.5%, the correct segment rate is 98.5%, the correct number identification rate is 99.6%, the address name recognition rate is 99.3%, Hiragana recognition rate is 99.5%, with an average each image recognition identified as 1.08s. Experimental results show that the algorithm improved and presented by my study of Japanese license plate recognition can solve a number of challenges including location of the license plate, character segmentation and recognition in complex environment and can meet the requirements of the product of Japanese license plate recognition system.